Machine learning based search method, system, and storage medium

Abstract

The present application relates to the technical field of intelligent search technology, and specifically relates to a search method and system based on machine learning and a storage medium, comprising the following steps: performing missing name target resolution on an original search request to generate a default reference hypothesis corresponding to the original search request; triggering non-target word search in multiple candidate information sources to extract a group of accompanying trace segments corresponding to the default reference hypothesis; identifying a default reference closed loop that collectively points to the same implicit target across information sources and generating a closed loop search expression based on the default reference closed loop; and outputting a target result sequence consistent with the default reference closed loop. The present application takes cross-source consistency as a core criterion, effectively eliminates accidental matches that only exist in a single information source, and strengthens results that have multi-source support, thereby reducing the rate of false matches.

Description

Technical Field

[0001] This invention relates to the field of intelligent search technology, and in particular to search methods, systems and storage media based on machine learning. Background Technology

[0002] In multi-source information collaborative processing and retrieval applications in complex environments, search systems typically rely on the explicit name, alias, or semantically approximate expression of the target object for matching to locate and extract target information. In existing technologies, whether it's keyword-based retrieval methods or vector-representation-based semantic retrieval methods, they are essentially based on the premise that the target can be explicitly expressed, meaning that the retrieval object exists in the data in the form of a stable name, standard number, or expandable alias.

[0003] However, in practical applications, especially in collaborative processing scenarios involving highly sensitive information, target objects often do not exist in a stable and complete name form. Due to various factors such as hierarchical management, stage switching, number rotation, summary transcription, and citation pruning during information generation, transmission, and recording, the expression of target objects can change, and they may even be intentionally omitted or replaced in some records, gradually forming a record form of "nameless subject, but traces present." Specifically, the target object does not appear directly in the text by its actual name, but rather leaves traces indirectly through approval succession relationships, version evolution relationships, temporal connections, spatial mapping relationships, code fragments, or gaps in related content.

[0004] In this context, existing retrieval methods face significant limitations. On the one hand, when the target name is cropped, replaced, or temporarily hidden, traditional keyword- or alias-based retrieval methods struggle to effectively capture relevant content. On the other hand, even vector retrieval or semantic recall methods often only capture superficially semantically similar text, failing to identify implicit information that, while not directly expressing the target, is actually related, easily generating numerous superficially relevant but actually unrelated search results. This type of problem is more subtle than the common synonym matching bias problem and is particularly prominent in multi-source collaborative scenarios. Summary of the Invention

[0005] This invention provides a machine learning-based search method, system, and storage medium that can break through the dependence on explicit target representations and no longer use the target name as the sole retrieval entry point. Instead, it mines indirect traces related to the target from multi-source records, and identifies evidence structures that commonly point to the same implicit target by modeling and determining the relationships between these traces, thereby achieving effective retrieval and reconstruction of unnamed targets.

[0006] Machine learning-based search methods include the following steps: S1. Perform missing name target parsing on the original search request to generate a default reference hypothesis corresponding to the original search request; S2. Based on the default reference hypothesis, trigger a non-target word search in multiple candidate information sources and extract the accompanying trace fragment group corresponding to the default reference hypothesis; S3. Input the accompanying trace fragment group into the machine learning closed loop recognition model to identify the default referential closed loop that points to the same implicit target across information sources, and generate a closed loop search expression based on the default referential closed loop. S4. Perform closed-loop verification search on the candidate information source using the closed-loop search expression, and output the target result sequence that is consistent with the default referential closed loop.

[0007] Optionally, the missing name target parsing includes performing syntactic analysis and semantic role labeling on the original retrieval request to identify the missing reference position in the original retrieval request and the context keywords associated with the missing reference position; based on a preset domain knowledge graph, matching the context keywords with entity nodes in the domain knowledge graph, and inferring the candidate hidden target entity corresponding to the missing reference position based on the association path between the matched entity nodes.

[0008] Optionally, the candidate implicit target entities are scored with confidence, and the candidate implicit target entities whose confidence scores meet a preset threshold are selected as the default referents. Combined with the context keywords, a default referential hypothesis is generated that includes the relationship between the default referent and the context keywords.

[0009] Optionally, the non-target search term set includes the context keywords and the derived feature words of the default referent target on the preset association dimension, and explicitly excludes the identifier words of the default referent target itself; the non-target search term set is extracted from the default referent assumption, and the non-target search term set is input into the multiple candidate information sources to perform parallel search, and the original search results page returned by each candidate information source is obtained; the original search results page is subjected to text extraction and fragment segmentation to generate multiple candidate text fragments; Based on the association between the default referential target and the context keywords in the default referential hypothesis, text fragments that co-occur with the association are selected from the multiple candidate text fragments, and the selected text fragments are deduplicated and aggregated to generate a group of accompanying trace fragments corresponding to the default referential hypothesis.

[0010] Optionally, after the accompanying trace fragment group is vectorized and encoded, multiple fragment feature vectors are generated; the multiple fragment feature vectors are input into the machine learning closed loop recognition model, and the semantic association strength between fragment feature vectors of different information sources is calculated through the attention mechanism of the machine learning closed loop recognition model, and the fragment feature vectors are clustered according to the semantic association strength to form multiple candidate closed loop clusters.

[0011] Optionally, for each candidate closed-loop cluster, the implicit target entity that is commonly pointed to by cross-information source fragments in the candidate closed-loop cluster is extracted, and the association path between the implicit target entity and the context keywords in the default referencing hypothesis is constructed. If the association path forms a closed semantic loop, the candidate closed-loop cluster is determined to be a default referencing closed loop. The target entity identifier and the key related words on the association path are extracted from the default referencing closed loop, and the target entity identifier and the key related words are combined according to a preset query syntax to generate a closed-loop search expression.

[0012] Optionally, the closed-loop search expression is decomposed into verification target words and verification related words. The verification target words are derived from the target entity identifier words in the default referencing closed loop, and the verification related words are derived from the key related words in the default referencing closed loop. The verification target word and the verification related word are combined into a verification query, and a verification search is performed in the multiple candidate information sources to obtain the verification search results page returned by each candidate information source. The verification search results page is parsed to extract verification result items that contain the verification target word and also contain the verification related word or its synonym, and a candidate result set is constructed.

[0013] Optionally, the semantic matching degree between each verification result item in the candidate result set and the default referential loop is calculated, verification result items with matching degrees lower than a preset threshold are filtered out, and the remaining verification result items are sorted in descending order of matching degree to generate a target result sequence consistent with the default referential loop.

[0014] A machine learning-based search system, used to implement the above search methods, includes the following modules: Missing Name Target Parsing Module: Performs missing name target parsing on the original search request and generates a default reference hypothesis corresponding to the original search request; Non-target search and trace extraction module: Based on the default reference hypothesis, trigger non-target word search in multiple candidate information sources to extract the accompanying trace fragment group corresponding to the default reference hypothesis; The closed-loop identification and expression generation module inputs the accompanying trace fragment group into the machine learning closed-loop identification model, identifies the default referential closed loops that point to the same implicit target across information sources, and generates a closed-loop search expression based on the default referential closed loops. The closed-loop verification and result output module is used to perform closed-loop verification search on the candidate information source using the closed-loop search expression, and output the target result sequence that is consistent with the default referent closed loop.

[0015] A computer-readable storage medium storing a computer program that can be loaded by a processor and execute the methods described above.

[0016] The beneficial effects of this invention are: This invention introduces a technical link of default referencing hypothesis—accompanying trace fragment group—default referencing closed loop. Instead of relying on target name, alias, or direct semantic matching, it starts from contextual keywords, combines knowledge graph association paths and multi-source fragment co-occurrence relationships to perform reverse inference and closed-loop confirmation of the implicit target. Compared to existing retrieval methods that rely on explicit naming or semantic similarity, this invention can still recover the target's location through multi-source indirect evidence even when the target's expression is truncated, replaced with a different name, or it doesn't appear at all. This solves the problem of having information but being unable to be retrieved, improving retrieval capabilities and applicability in complex scenarios.

[0017] This invention uses a machine learning closed-loop recognition model to jointly model the feature vectors of fragments from different information sources, calculates the semantic association strength between fragments using an attention mechanism, and further identifies whether a closed semantic loop of target-relationship-context is formed through path construction and closed-loop judgment function. This mechanism breaks through the limitations of single-source ranking or simple similarity judgment in existing technologies, and uses cross-source consistency as the core criterion, which effectively eliminates accidental matches that only exist in a single information source, while strengthening the results that are truly supported by multiple sources, thereby reducing the false matching rate.

[0018] This invention, building upon closed-loop recognition, further introduces a verification search mechanism. It decomposes the closed-loop search expression into verification target words and verification related words, and uses a dual-hit rule and matching degree scoring function to perform secondary screening and ranking of candidate results. This method not only ensures that the results clearly point to the target but also requires them to be in the correct semantic context, thus avoiding the problems of mentioning the target in the wrong context or having a similar context but an incorrect target. Simultaneously, through quantified matching scoring and threshold control, the result screening process possesses clear interpretability and adjustability, ensuring both accuracy and recall, thereby forming a stable and reliable target result sequence output. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only for this invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a schematic diagram of the method flow according to an embodiment of the present invention; Figure 2 This is a logic block diagram of an embodiment of the present invention. Detailed Implementation

[0021] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. For some well-known technologies, those skilled in the art may also use other alternative methods to implement the invention. Moreover, the accompanying drawings are only for more specific description of the embodiments and are not intended to specifically limit the present invention.

[0022] like Figures 1-2 As shown, the machine learning-based search method includes the following steps: S1. Perform missing name target parsing on the original search request to generate a default reference hypothesis corresponding to the original search request.

[0023] S11, after performing syntactic analysis and semantic role labeling on the original retrieval request, the original retrieval request is represented as a word sequence: ; Identify the set of missing positions using semantic role labeling functions: ; in, This represents the word sequence of the original search request. Indicates the first 1 term, Indicates the first The index refers to the missing position, the first term in the original search request word sequence. The original search request is represented as an ordered sequence of words corresponding to each position. Indicates the first One location index was identified as referring to a missing location. In other words, NULL ROLE indicates that the actual word or placeholder at that position is missing a marker that lacks a clear semantic reference. This represents a semantic role labeling function, used to identify the semantic function of each word in the original search request. Its core purpose is to determine whether the word plays a clear semantic role in the sentence, or whether there is a semantic gap.

[0024] In practice, syntactic dependency analysis is performed on the original retrieval request to obtain the dependency relationships between each word and its context, including subject-predicate, verb-object, and modification relationships, thus forming a syntactic structure tree. Secondly, based on the syntactic structure, an encoder is used to learn the context representation of the sentence, and then sequence labeling or pointer mechanisms are used to identify the semantic role type corresponding to each word, achieving role labeling of predicates and their related components in the sentence. During the labeling process, each word is assigned one of the following types of tags: core semantic role, auxiliary semantic role, non-semantic role, or missing semantic role. The missing semantic role tag indicates that the word's position should have a certain semantic role in the syntactic structure, but the corresponding entity or object is not explicitly given in the current expression. Finally, the output of the semantic role labeling function is to return the semantic role label for each word.

[0025] After obtaining the semantic role annotation results for each word, it is necessary to further identify which positions belong to the missing referents. The specific identification logic is as follows: First, iterate through all predicates in the sentence. For each predicate, check whether its core role in the semantic framework is complete. For example, an action should usually have an "executor" and an "object of action". If only the action description appears in the sentence, but the corresponding object or subject is missing, it means that there is a semantic missing at that position.

[0026] Secondly, semantic locations that should exist but do not appear are mapped back to specific locations in the original word sequence. These locations may correspond to omitted object positions, words that have been replaced with vague references, or implicit locations that exist semantically but are not explicitly stated in the text.

[0027] Then, the words corresponding to these positions are marked. When the semantic role labeling function outputs NULL ROLE, the position is considered to be a missing reference position. Finally, all position indices that meet this condition are collected to form a set of missing reference positions. In other words, a missing reference position is not that there is no word, but that there is a word but lacks a clear semantic reference, or that the position should have a clear object but is omitted or ambiguous.

[0028] S12, for each vacant reference position Extract within its neighborhood window: ;in, Indicates the relationship with the first Each refers to the context corresponding to the vacant position. This indicates the radius of the context window, used to limit the scope of keyword extraction. When the sentence is short, the window can be appropriately reduced; when the sentence is long, the window can be appropriately expanded. Short sentences extract 2-3 words, and long sentences extract 3-6 words. This represents the word corresponding to the j-th position in the original search request. When the semantics of a certain position in a sentence are incomplete, words that can provide clues are extracted from a small segment of text before and after the position itself, i.e., contextual keywords. These contextual keywords are the basis for matching entities and inferring implicit targets later.

[0029] S13, Entity Node Matching: Based on the domain knowledge graph, context keywords are mapped to a set of entity nodes, turning context keywords into entities in the knowledge graph, represented as: ;in, Represents the set of context keywords The set of entity nodes obtained by matching For entity nodes in a knowledge graph, Keywords With entity nodes The semantic similarity function is based on the cosine similarity function derived from vector representation. This represents the entity matching threshold, with a value of 0.75. This value corresponds to a semantic similarity in the medium to high similarity range, which can effectively filter out semantically deviating entities while retaining the matching results of synonymous and approximate expressions, achieving a balance between precision and recall. In vector matching based on cosine similarity, 0.75 is usually used as a stable dividing point.

[0030] Domain knowledge graphs are web databases that represent entities and relationships in a structured way, and consist of three parts: Entity nodes include objects, events, numbers, and time segments; Relationship edges include those that belong to, are associated with, are inherited from, or occur in; Attributes: Used to describe the characteristic information of an entity.

[0031] The domain knowledge graph is constructed as follows: 1. Extract information from historical records, structured databases, and text archives.

[0032] 2. Entity extraction: Identify key objects, such as names, numbers, and time points.

[0033] 3. Relationship Extraction: Identify the relationships between entities, such as who is associated with whom, who evolved from whom, and who belongs to which stage.

[0034] 4. Standardize different expressions, including unifying different names for the same object and standardizing numbering formats.

[0035] 5. Graph structure construction: Entities are used as nodes and relationships are used as edges to form a graph structure, which is dynamically supplemented and corrected based on new data.

[0036] S14, In the domain knowledge graph, calculate the set of association paths between entity nodes: ; And calculate the path association strength: ; in, Represents the set of associated paths between entity nodes. Represents entity nodes and A connection path between them, This indicates whether a path exists between two entities, determined by performing a graph traversal search within the knowledge graph, starting from the entity... Starting from this point, we traverse along the relationship edges in the knowledge graph. If the entity can be reached within the set maximum search depth... If a path exists between the two entities, then it is considered that a path exists; otherwise, it is considered that no path exists. In other words, moving from one entity to another along a relationship chain is considered to involve a path. Indicates the path association strength. Indicates path length. For the first in the path The weight of an edge represents the importance of each relationship in the path. It is set based on the relationship type, with different relationships assigned different weights. Directly corresponding strong semantic relationships are given high weights, while indirectly related weak semantic relationships are given low weights.

[0037] S15, Based on path association strength, perform aggregation inference on candidate hidden target entities: ; in, Indicates the first Each refers to the set of candidate implicit target entities corresponding to the vacant position. Indicates the candidate implicit target entity. Representing entities With entity Path association strength, This represents the candidate screening threshold, with a value of 0.6 × maximum association strength. The actual analysis here is not about whether there are related entities, but which entity is the intersection point that multiple clues point to, that is, upgrading from relevance to indicativeness.

[0038] Confidence scores are assigned to candidate latent target entities: ; in, This indicates the confidence level of the candidate implicit target entity. The weighting coefficient is set to 0.6. The core objective of this scheme is to recover the implicit referent from unnamed target scenarios. Relying solely on surface textual similarity is easily influenced by vague, abbreviated, and substitute expressions. Therefore, the weighting of structural association evidence should be appropriately increased to ensure that the candidate implicit target entity is primarily supported by the fact that multiple entities jointly point to it. At the same time, the weighting coefficient should not be set too high, otherwise it will weaken the constraint of contextual keywords on the current retrieval context, easily leading to misjudgment of structurally centered entities in the graph as targets. A weighting of 0.6 indicates that structural association evidence is primary, supplemented by semantic fit evidence, which better aligns with the technical objectives of this scheme. This represents the overall semantic similarity between the entity and the set of context keywords. The value represents the size of the entity node set. This confidence score doesn't solely rely on the similarity between the candidate hidden target entity and the context keywords, nor does it solely rely on the strength of the connection between the candidate hidden target entity and existing entity nodes. Instead, it combines these two types of evidence. The former reflects structural association evidence, i.e., whether the candidate hidden target entity is located at the convergence point of multiple entity nodes; the latter reflects semantic fit evidence, i.e., whether the candidate hidden target entity is semantically consistent with the context keywords. The former is better at identifying entities that are mutually referencing in terms of relationship, while the latter is better at identifying entities that are truly close to the current search intent in terms of text. Therefore, the weighted combination of these two factors avoids structural relevance but semantic deviation caused by relying solely on path relationships, and also avoids textual similarity but lack of graph support caused by relying solely on semantic similarity. Thus, this formula simultaneously considers structural credibility and semantic consistency.

[0039] Candidate implicit target entities that meet the threshold condition are selected as the default referent targets: ;in, The default target set is used. The confidence threshold is set to 0.7. The confidence score is already a comprehensive result of the fusion of structural association evidence and semantic fit evidence. Therefore, the threshold should be set in the medium to high range to ensure that the candidate latent target entities that are ultimately retained have both good graph support and semantic consistency. If the threshold is too low, many entities with weak association and weak semantic matching will also be included in the default referent, increasing the interference when extracting subsequent accompanying trace fragment groups. If the threshold is too high, it is easy to remove some real targets with slightly lower semantic scores due to expression pruning or incomplete context, but which actually have strong structural references. Therefore, setting it to 0.7 can achieve a good balance between accuracy and retaining potential real targets.

[0040] S16, the default pronoun hypothesis is generated as follows: ;in, Indicates the first The default reference assumptions for each missing reference position This indicates the default target. For a set of context keywords, This is a set of entity association paths.

[0041] S2. Based on the default reference hypothesis, trigger a non-target word search in multiple candidate information sources to extract the accompanying trace fragment group corresponding to the default reference hypothesis.

[0042] In S1, the default referential hypothesis was obtained, including two key elements: the possible target (the default referential target) and the contextual keywords describing this target. However, in S2, the target itself is deliberately avoided in the search, as the target name may be hidden, replaced, abbreviated, or not appear at all. Therefore, the contextual keywords are extracted, and some indirect feature words are derived from the target itself. These two types of words are then combined as search input, while the target's name is explicitly removed. This makes it easier to find content that does not directly mention the target but describes it. Non-target search terms are then searched across multiple information sources, yielding many web pages or text results. The main text is then extracted from the results, and long texts are cut into smaller segments. Keyword filtering is then performed, retaining only segments that simultaneously demonstrate the target's relevance and contextual information. In other words, keywords are not accepted randomly, but must match the relationship characteristics that the segment actually describes the implicit target. Finally, duplicate content is removed, and related segments are grouped together to form so-called accompanying trace segment groups, which are essentially sets of evidence that do not directly state the target but indirectly describe it. Specifically: S21, extract the set of non-target search terms from the default referential hypothesis, defined as: ; And it satisfies: ; in, For the first The default referential hypothesis corresponds to a set of non-target search terms, representing a group of words used to replace the search without using the target name. It consists of context keywords and derived feature words. Context keywords are clues already present in the original sentence, while derived feature words are indirect features inferred from the target. Their essential function is to bypass the target itself and search for descriptions of its environment and behavior. For a set of context keywords, This refers to the set of derived feature words of the default referent target in the preset association dimension. The default target set is used. Ø represents the set of identifiers for the default target, where Ø is an empty set, indicating that identifiers for the default target itself are explicitly excluded. The set of identifiers for the default target consists of words that, once they appear, can directly point to the target, such as formal names, numbers, unique codes, or explicit aliases. Once these words are used in a search, it is equivalent to directly exposing the target. However, the core of this solution is precisely to avoid directly searching for the target, so all these identifiers must be excluded.

[0043] Predefined association dimensions are a set of indirect observation angles used to describe a target, rather than directly using default names or identifiers. In other words, it doesn't directly ask what the target is, but rather describes the information that typically accompanies the target from several stable external associations. These observation angles are the predefined association dimensions. They need to be pre-organized based on recording habits, retrieval habits, and target concealment methods in a specific domain. In this invention, the predefined association dimensions are set as time association dimensions, spatial association dimensions, behavioral association dimensions, relational association dimensions, attribute association dimensions, and evolutionary association dimensions. The time association dimension describes the target's frequently occurring time characteristics, stage characteristics, or chronological clues; the spatial association dimension describes the target's location, region, deployment point, or spatial mapping relationship; the behavioral association dimension describes the target's related actions, processes, triggering events, or processing behaviors; the relational association dimension describes the target's subordinate, collaborative, referencing, inheriting, or mapping relationships with other objects; the attribute association dimension describes the target's category, status, purpose, level, feature tags, and other static information; and the evolutionary association dimension describes the target's substitution, change, continuation, splitting, merging, or version evolution clues at different stages. The reason for pre-setting these dimensions is that the default referent often doesn't appear by its actual name in many texts, but it will consistently leave accompanying traces through these dimensions. Therefore, by first fixing which angles are most likely to leave traces, we can then generate non-target search terms in a targeted manner. In terms of the pre-setting method, domain experts first summarize which related dimensions best represent the implicit existence of the target based on historical corpora, business documents, naming conventions, and record templates. Then, they verify whether these dimensions are effective in recovering the target through historical search results, retaining dimensions with strong discriminative ability, high stability, and low noise as pre-set related dimensions.

[0044] The default referent is a set of derived feature words along preset association dimensions. After determining the preset association dimensions, the name of the default referent itself is no longer used directly. Instead, terms that can indirectly represent the target are extracted along each preset association dimension, and then these terms are aggregated to form a set. First, with the default referent as the center, descriptive information with a clear association with the target is searched in the domain knowledge graph, historical document library, structured record table, terminology mapping table, or existing annotated corpus. Then, this descriptive information is classified and organized according to the preset association dimensions. In the time-related dimension, extract the stage name, time segment, periodic words, and time sequence nodes corresponding to the target; in the spatial-related dimension, extract the location names, area markers, deployment locations, and coordinate segments related to the target; in the behavioral-related dimension, extract the action words, process words, disposal words, and invocation words that appear together with the target; in the relational-related dimension, extract the names or relational expressions of related objects that have reference, inheritance, subordination, matching, or mapping relationships with the target; in the attribute-related dimension, extract category words, status words, feature tags, and usage tags; and in the evolution-related dimension, extract historical alternative names, version number fragments, stage aliases, and change trace words.

[0045] In step S22, the resulting set of non-target search terms is constructed around the indirect features of the target. Next, search requests are simultaneously initiated from multiple information sources, including different databases, document systems, and indexes. During execution, each candidate information source independently receives the same set of non-target search terms and returns its respective search results page. These result pages are essentially the collection of all raw content that the information source considers relevant to these search terms. Because execution is parallel, multiple information sources do not affect each other, but ultimately form a more comprehensive pool of result sources. The final result is not the processed result, but the raw search results page—the initial data before cleaning, segmentation, and filtering—preparing for subsequent text extraction and fragment filtering. Candidate information sources refer to the set of all data sources that can be invoked to perform search operations, including structured databases, document repositories, log systems, historical record repositories, index engines, or distributed data nodes. Each information source stores different types or focuses of data, and can return content related to the search terms from different perspectives.

[0046] Specifically, the non-target search term set is input into multiple candidate information sources to perform parallel search, defined as: ;in, Indicates the first A collection of original search results pages returned by each candidate information source. For the first One candidate information source This is a set of non-target search terms. This indicates a search execution. The specific execution steps for a search execution are as follows: Standardize the non-target search term set, including unifying word segmentation, unifying synonyms, and standardizing format, so that it can be adapted to the query interface of different information sources; Based on the query rules of each candidate information source, the standardized search term set is converted into the corresponding information source query expression, such as constructing keyword combinations, Boolean expressions, or search statements; The transformed query expression is sent to the corresponding information source interface, triggering the retrieval mechanism within that information source; The information source internally performs an index matching process, retrieves records related to the query expression from its data set, generates a result list according to its internal sorting rules, and returns the results in the form of pages or records, including titles, summaries, body text or related field information. The results returned by each information source are encapsulated in a unified format to form the original search results page set corresponding to that information source.

[0047] S23. The original search results page is usually quite cluttered, containing not only the main text but also irrelevant information such as navigation bars, advertisements, tags, links, headers, and footers. Furthermore, the content is often in whole paragraphs or even entire pages, which is not conducive to subsequent detailed analysis. Therefore, we extract the main text and segment it into fragments from the original search results page to generate a set of candidate text fragments: ; in, Represents a set of candidate text fragments. This indicates text extraction, used to extract the main text content from a page. This indicates segmentation, used to divide the main text into multiple text segments. This represents the collection of original search results pages.

[0048] 1. The specific steps for extracting the main text are as follows: The original search results page is structurally analyzed to identify the different areas of the page, including the title area, body text area, navigation area, advertising area, and footer area. Based on the page tag structure or text density characteristics, each region is initially classified. The main text region usually has the characteristics of high continuous text density, concentrated punctuation, and semantic coherence. Remove content outside the main text area, including navigation menus, recommended links, advertising information, script code, and duplicate template content; The remaining text is cleaned, including removing HTML tags, removing special characters, standardizing the encoding format, and correcting abnormal line breaks; The text undergoes a coherence check, and incorrectly segmented sentences are reassembled to ensure semantic integrity. Finally, the structured main text content is output as input for subsequent segmentation.

[0049] 2. The specific segmentation is as follows: Perform basic sentence segmentation on the main text, dividing the text into initial sentence units based on punctuation marks; Based on sentence length and semantic integrity, excessively short sentences are merged, and excessively long sentences are split, so that each segment maintains a moderate length; Based on semantic continuity, adjacent and semantically closely related sentences are grouped into a segment, such as the same event description or the same logical unit; Each fragment undergoes boundary correction to avoid erroneous truncation in key entity or relation representations. Fragments are numbered or identified to form an indexable, structured fragment set. Finally, all text fragments are output to form a candidate text fragment set.

[0050] S24. As mentioned above, all search results are broken down into multiple small text fragments. There are many fragments, some of which just happen to contain keywords, and some are only semantically similar but unrelated. Only a portion of the fragments truly indirectly describe the default referent. Therefore, it is necessary to filter out irrelevant fragments and keep only those that meet two conditions: Condition 1: The fragment must be related to the default referent, that is, the fragment must contain an expression that reflects the target; Condition 2: The fragment must also contain contextual keywords, that is, contextual information related to the current search request must appear simultaneously. Only fragments that meet both conditions are considered not to be accidental matches, but to be valid evidence that truly points to the implicit target.

[0051] Therefore, based on the association between the default referent target and context keywords in the default referent hypothesis, text fragments that meet the co-occurrence condition are selected: This formula represents selecting, from all candidate text fragments, those fragments that contain at least one set of target-related expressions and contextual keywords. For the filtered collection of text fragments, For candidate text fragments, The default target set is used. For a set of context keywords, Co-occurrence determination is indicated when the text fragment... Existence and Target Related expressions and keywords appearing simultaneously The value is 1 if it is true, and 0 otherwise.

[0052] S25, For the filtered set of text fragments The process involves deduplication, merging segments with identical content or highly similar expressions into a single group to avoid repetition of information affecting subsequent judgments. Next, aggregation is performed, grouping scattered but related segments together to form a structured evidence set. Centered on the default referent target in the default referent hypothesis, all segments are grouped according to their strength of association with that target. Then, further, segments are divided into subgroups based on the type or association of contextual keywords; for example, segments with the same time feature are grouped together, and segments with the same behavioral feature are grouped together. Within each group, segments are ordered according to chronological order, logical order, or semantic similarity to create continuous or interpretable relationships between them. Finally, these grouping results are organized into a cohesive structure, allowing different segments to complement each other and all pointing to the same implicit target, resulting in the accompanying trace segment group. .

[0053] S3. Input the accompanying trace fragment group into the machine learning closed loop recognition model, identify the default referential closed loop that points to the same implicit target across information sources, and generate a closed loop search expression based on the default referential closed loop.

[0054] S31, each text fragment in the accompanying trace fragment group Vectorization encoding is performed, converting each text segment into a numerical vector that can express its overall semantics, generating multiple segment feature vectors. , Indicates the first in the group of accompanying trace fragments A text fragment.

[0055] S32, input multiple fragment feature vectors into the machine learning closed-loop recognition model, and calculate the semantic association strength between feature vectors of different information source fragments through an attention mechanism: ;in, Representing fragments With fragments The strength of semantic association between them Represents the feature vector of a segment. This represents the query matrix and key matrix in the attention mechanism. The vector dimension is used for scaling. Semantic association strength characterizes the probability or strength of whether two segments refer to the same latent object. The machine learning closed-loop recognition model adopts a multi-segment relationship modeling structure based on an attention mechanism. Its core objective is to jointly model the feature vectors of segments from different information sources, thereby identifying whether there is a relationship between these segments that points to the same latent target. The model structure includes the following: Input representation layer: Receives multiple segment feature vectors obtained by vectorizing the accompanying trace segment group. Each segment feature vector corresponds to a text segment, and each segment is also appended with an information source identifier code to distinguish segments from different sources, enabling the model to explicitly perceive the "cross-information source" relationship features.

[0056] Feature mapping layer: The input fragment feature vector is mapped to three types of representations: query representation, key representation and value representation. This mapping is achieved through linear transformation, enabling the model to understand the semantic features of the same fragment from different perspectives, and providing a foundation for subsequent relation calculation.

[0057] Multi-head attention relationship modeling layer: In this layer, each segment establishes relationships with all other segments. Semantic associations in different subspaces are computed in parallel through a multi-head attention mechanism, thereby capturing multi-dimensional segment relationships. Some attention heads focus on capturing semantic similarity, some on capturing relationship consistency, and others on capturing contextual complementarity. The fusion of these multi-head mechanisms yields a more stable representation of segment relationships.

[0058] Cross-source relationship enhancement layer: This layer highlights the strength of the association between different information sources. By introducing information source difference constraints, segments from different information sources that are highly semantically consistent are given higher weights, while segments from the same information source that are highly repetitive are suppressed, thereby enhancing the ability to distinguish cross-source consistency.

[0059] Clustering Representation Layer: The fragment representations after relation modeling are re-encoded so that each fragment not only contains its own semantic information, but also incorporates the association information with other fragments, thus forming a high-order representation suitable for clustering. This representation can be used to aggregate fragments with strong semantic associations into candidate closed-loop clusters.

[0060] Output layer: Outputs the semantic association strength matrix between segments and the enhanced segment representation, providing the input basis for subsequent loop closure determination and clustering processes.

[0061] The machine learning-based closed-loop recognition model is trained using supervised or semi-supervised learning. Its core objective is to enable the model to correctly distinguish between sets of segments belonging to the same latent target and sets of semantically similar segments that do not belong to the same target. First, training data is constructed. The training data consists of multiple sets of segments, each containing several text segments from different information sources, labeled with whether they point to the same latent target. Specifically, there are two categories: positive samples and negative samples. Positive samples are multiple segments that genuinely describe the same target, while negative samples are segments that, although semantically similar, point to different targets. Next, the segments from each training sample are input into the model to obtain the semantic association strength between the segments. For positive samples, the model should output a high association strength; for negative samples, the model should output a low association strength. The model is then optimized by constructing a loss function. The loss function can be designed based on the difference between the association strength and the true label, allowing the model to gradually learn to distinguish between true and false associations during training. During training, the model parameters are continuously adjusted to make segments with the same target more semantically close, while segments with different targets are more semantically separated. Meanwhile, cross-information source constraints can be introduced during the training process, which requires the model to prioritize strengthening the association between semantically consistent segments from different information sources and reduce the influence of repeated segments within the same information source, thereby enhancing the model's ability to identify cross-source closed loops. Finally, after multiple rounds of training, the model can automatically output a stable semantic association strength when any set of segment feature vectors is input, and provide a reliable basis for subsequent closed loop identification.

[0062] Based on semantic association strength, the fragment feature vectors are clustered to form candidate closed-loop clusters: ; in, Indicates the first A candidate closed-loop cluster, Representing fragments The strength of semantic association with other segments, This represents the clustering threshold, used to filter highly correlated segments. When the semantic correlation strength has been normalized to the range of 0 to 1, the value is 0.65. This threshold is located in the medium to high correlation range and can effectively distinguish between true semantic correlation and superficial similarity.

[0063] The execution process of the machine learning closed-loop recognition model in this invention is as follows: Step 1: Prepare the fragment vectors. Each text fragment has been transformed into a vector.

[0064] The second step is to calculate the attention mechanism. For any two segments, segment A is used to query segment B to see if they are related. Specifically, each segment generates a query perspective and a query feature. Then, the matching degree between the query perspective and the query feature is calculated. The matching degree indicates whether segment A thinks segment B is related to itself.

[0065] The third step is to obtain the semantic association strength. After calculating the relationship between all segments, a relation matrix will be obtained. That is, each pair of segments has an association strength value. The larger the value, the more likely the two segments are to describe the same implicit target.

[0066] Step 4: Cross-information source reinforcement. Since the input fragments come from different information sources, if there is a strong correlation between fragments from different sources, it means that it is not accidental, but a stable direction. This is the key to the present invention: using cross-source consistency to identify the real target.

[0067] Step 5: Used for subsequent clustering. The semantic association strength mentioned above will be used as a basis to group together fragments with strong associations to form candidate closed-loop clusters.

[0068] S33, for each candidate closed-loop cluster, extract the hidden target entity that is commonly pointed to by cross-information source fragments: ;in, Indicates the first The implicit target entities corresponding to each candidate closed-loop cluster Representing entities With fragment vector The association function. Based on keyword hit rate calculation, the set of keywords corresponding to the entity is defined as... The set of words in the fragment is The keyword hit rate can then be calculated as follows: ;in, For entities Keyword set, Representing fragments a set of words, This indicates the number of keywords that were hit. This indicates the total number of entity keywords.

[0069] Constructing the association path between the implicit target entity and the context keywords: ;in, Represents a set of associated paths. Indicates a related path. Represents a set of context keywords. This function represents a path search. It determines whether there is a connection path between entity e and keyword k in the knowledge graph and finds the path. First, keyword k is mapped to entity nodes in the knowledge graph. Starting from entity e, the search is performed in the knowledge graph. The maximum search depth is set to 2 to 3 levels. During the search, if the node corresponding to k is reached within the limited depth, it is considered that a path exists. At the same time, the nodes and relationships on the path are recorded. If multiple paths exist, the shortest path or the path with the highest association strength is retained. Finally, one or more relationship paths from entity to keyword are output.

[0070] If the associated paths form a closed semantic loop, it is determined to be a default referencing closed loop: ; in, This represents the closed-loop determination function.

[0071] The first step is to construct a relationship graph: forming a subgraph structure by combining the implicit target entity, context keywords, and intermediate nodes in all paths; The second step is to determine connectivity: check whether all keyword nodes can be connected to the target entity through a path. If any are disconnected, the condition is not met. The third step is to determine the closure: check for the existence of loops or cross connections, including shared nodes between different paths or the formation of loop structures between paths. This indicates that the information is not derived in a one-way manner, but rather mutually supportive. Step 4: Check whether the path relationships are semantically consistent, including whether there are conflicting relationships or logical breaks. The judgment rule is: if all keywords are reachable, there are intersections or loops between paths, and there are no obvious semantic conflicts, then it is judged as a closed loop and output 1; otherwise, it is judged as a non-closed loop and output 0.

[0072] S34, extract the target entity identifier and key related words on the associated path from the default referential closed loop to generate the closed loop search expression: ;in, This represents a closed-loop search expression. A set of identifiers representing implicit target entities. This represents the set of key related words along the association path. This indicates query combination, used to generate expressions according to a preset query syntax. That is, it concatenates target keywords and related terms into an executable search statement according to rules. The specific steps are as follows: First, select one or more of the most representative words from the target entity identifier set as core search terms. Then, select several words that reflect contextual relationships from the key related words set on the association path as auxiliary search terms. Next, combine the core search terms and auxiliary search terms according to the preset query syntax, using a combination method that requires simultaneous appearance, and connect the core terms and auxiliary terms with a logical AND. For multiple auxiliary terms, OR can be used as needed to enhance recall capability. Finally, the combined expression is standardized to generate a closed-loop search expression with a unified format.

[0073] S4. Perform closed-loop verification search on the candidate information source using the closed-loop search expression, and output the target result sequence that is consistent with the default referential closed loop.

[0074] In S41, the default referential loop was obtained through loop closure identification in S3. This loop has identified a relatively credible implicit target and a series of related relationships surrounding that target. However, this information remains at the inference level, so further verification is needed. This involves re-initiating a search in a more explicit and controllable manner to confirm the validity of the loop. Therefore, the original loop search expression needs to be deconstructed and reconstructed. The original loop search expression is often a comprehensive expression, containing both target and relational information; direct use might introduce noise or instability. Therefore, it is broken down into two clearer components: words directly pointing to the target and words describing the target's context or relationships. These two types of words are then recombined into a verification query to initiate a more precise verification search. The target words ensure the search direction is correct, and the related words ensure the search context is correct. The combination of these two filters out results that only mention the target but have incorrect context or have similar contexts but incorrect targets.

[0075] Specifically, the closed-loop search expression is decomposed into verification target words and verification related words. Verification target words are directly derived from the identification information of implicit target entities, such as names, numbers, and aliases. Verification related words are derived from key relationship expressions in the association path, such as behavioral description words, relational words, and feature words. In other words, they are divided according to whether they directly point to the target. Those directly pointing to the target are classified as verification target words, while the rest, used to describe relationships or context, are classified as verification related words. A verification query is then constructed: ;in, Indicates the first The default reference refers to the verification query corresponding to the closed loop. This represents the set of target words for verification, derived from target entity identifier words. This represents the set of validation related terms, derived from key related terms. This represents a query combination function. The query combination function selects one or more core terms from the target term set as the main search term; selects several auxiliary terms from the related term set as context constraint terms; concatenates the main search term and auxiliary terms according to logical combination rules, where the main search term and auxiliary terms must be combined in a way that satisfies both conditions simultaneously, while multiple auxiliary terms can be combined in a way that satisfies at least one condition to improve recall; finally, the combined result is formatted to generate a unified validation query, essentially constructing a search expression that both targets and restricts context.

[0076] S42, perform a verification search on the verification query across multiple candidate information sources to obtain the verification search results page: ;in. Indicates the first The closed loop is in the first... A collection of verified search results pages returned by each candidate information source. Indicates the first One candidate information source To validate the query, a validation query is a search expression with dual constraints, consisting of a validation target term and validation related terms. Its purpose is to ensure that the search results revolve around the target and appear in the correct semantic relationship or context, thereby effectively validating the implicit target. This indicates the search execution, which is the same as the search execution in S22. First, the verification query is normalized to adapt it to the query format of different information sources; the query is converted into a retrieval expression that can be executed by each information source; then it is sent in parallel to multiple candidate information sources to trigger the search; subsequently, each information source performs internal matching and sorting and returns the results page; finally, the returned results are uniformly encapsulated to form a set of verification search results for subsequent filtering.

[0077] S43. After completing the verification search, a batch of verification search results pages are obtained. Not all results are truly valuable. The goal is to filter out those results that can truly support both the existence of the target and the validity of the relationship, and build a more reliable candidate result set. First, each result item must explicitly mention the target itself, that is, it must hit the verification target word. This ensures that the result is indeed talking about the object you care about, and not some other similar object. Second, on the premise of meeting the target requirement, the result must also reflect the corresponding relational context, that is, hit the verification related words or their synonyms. This ensures that the result not only mentions the target, but also mentions it in the correct semantic context, rather than an irrelevant reference.

[0078] The verification search results page is parsed, and verification result items that meet the conditions are filtered to construct a candidate result set: This formula represents selecting, from all verified search results, result items that simultaneously contain the target word and related words or their synonyms, forming a candidate result set. Indicates the first The candidate result set corresponding to each closed loop For a single validation result item, Here is the hit function, representing the result item. Does it contain a set of words? For any word in the given set, the resulting item is segmented and normalized. Then, the word set is matched with the target word or related word set. If the resulting item contains at least one word that is the same as or semantically equivalent to the given word set (i.e., includes synonyms), it is considered a match, and the output is 1; otherwise, the output is 0. This represents the set of synonyms for the verification related words. This is a logical AND operation, indicating that two conditions must be met simultaneously.

[0079] S44, Calculate the semantic matching degree between each verification result item in the candidate result set and the default referential loop: ;in, Indicates the verification result item semantic matching degree This represents the weighting coefficient; a larger value indicates greater emphasis on whether the target is mentioned. The larger the value, the more emphasis is placed on whether the context is correct. This invention uses a value of 0.6, appropriately increasing the weight of verifying the target word hit, prioritizing ensuring that the result points to the correct target, while retaining a certain proportion of contextual constraints to avoid mentioning the target but with incorrect context. It's the weighted sum of these two parts, which can be understood as a comprehensive score combining the degree of target hit and the degree of context matching. Represents the result item a set of words, To verify the target word set, This represents the set of related words to be verified. This represents the set of synonyms for the verification related words. Indicates the number of elements in the set.

[0080] This indicates the percentage of target words that were matched in this result item, representing the proportion of all target words matched. It is the set of words in the result item. It verifies the target word set. The intersection of the two sets is the actual number of target words hit. Dividing this number by the total number of target words gives the hit rate. Essentially, it analyzes whether the result truly mentions the target and whether it mentions it sufficiently. This indicates how many related words or their synonyms were matched in the result item, representing the proportion of all related words. Here, a set of synonyms is introduced to avoid missed matches due to differences in expression.

[0081] S45, filter out validation result items with a matching degree lower than the threshold, and sort them to generate the target result sequence: ;in, Represents the target result sequence. The matching threshold is set to 0.7. The matching score itself is a comprehensive score obtained by weighting the target hit ratio and the context matching ratio. Therefore, it needs to be set in the medium to high range to ensure the quality of results. When the threshold is below 0.6, it is easy to retain a large number of results that are only partially matched or have incomplete context, resulting in increased noise. When the threshold is above 0.8, it may filter out some results that are incomplete but actually effective, reducing recall. Setting it to 0.7 achieves a good balance between result purity and coverage. This represents a sorting function that sorts data from highest to lowest matching degree. This is the candidate result set.

[0082] The present invention also includes a machine learning-based search system for implementing the above-described search method, comprising the following modules: Missing Name Target Parsing Module: Performs missing name target parsing on the original search request and generates a default reference hypothesis corresponding to the original search request; Non-target search and trace extraction module: Based on the default reference hypothesis, trigger non-target word search in multiple candidate information sources to extract the accompanying trace fragment group corresponding to the default reference hypothesis; The closed-loop identification and expression generation module inputs the accompanying trace fragment group into the machine learning closed-loop identification model, identifies the default referential closed loops that point to the same implicit target across information sources, and generates a closed-loop search expression based on the default referential closed loops. The closed-loop verification and result output module is used to perform closed-loop verification search on the candidate information source using the closed-loop search expression, and output the target result sequence that is consistent with the default referent closed loop.

[0083] The present invention further includes a computer-readable storage medium storing a computer program that can be loaded by a processor and execute the methods described above.

[0084] This invention encompasses any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of this invention. To provide the public with a thorough understanding of this invention, specific details are described in detail in the following preferred embodiments; however, those skilled in the art will fully understand the invention even without these details. Furthermore, to avoid unnecessary misunderstanding of the essence of this invention, well-known methods, processes, procedures, components, and circuits are not described in detail.

[0085] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A machine learning-based search method, characterized in that, Includes the following steps: S1. Perform missing name target parsing on the original search request to generate a default reference hypothesis corresponding to the original search request; S2. Based on the default reference hypothesis, trigger a non-target word search in multiple candidate information sources and extract the accompanying trace fragment group corresponding to the default reference hypothesis; S3. Input the accompanying trace fragment group into the machine learning closed loop recognition model to identify the default referential closed loop that points to the same implicit target across information sources, and generate a closed loop search expression based on the default referential closed loop. S4. Perform closed-loop verification search on the candidate information source using the closed-loop search expression, and output the target result sequence that is consistent with the default referential closed loop.

2. The machine learning-based search method according to claim 1, characterized in that, The missing name target parsing includes performing syntactic analysis and semantic role labeling on the original search request to identify the missing reference position in the original search request and the context keywords associated with the missing reference position; based on a preset domain knowledge graph, matching the context keywords with entity nodes in the domain knowledge graph, and inferring the candidate hidden target entity corresponding to the missing reference position based on the association path between the matched entity nodes.

3. The machine learning-based search method according to claim 2, characterized in that, The candidate implicit target entities are scored with confidence. Candidate implicit target entities whose confidence scores meet a preset threshold are selected as default referents. Combined with the context keywords, a default referential hypothesis is generated that includes the relationship between the default referent and the context keywords.

4. The machine learning-based search method according to claim 1, characterized in that, The non-target search term set includes the context keywords and the derived feature words of the default referent target on the preset association dimension, and explicitly excludes the identifier words of the default referent target itself; the non-target search term set is extracted from the default referent assumption, and the non-target search term set is input into the multiple candidate information sources to perform parallel search, and the original search results page returned by each candidate information source is obtained; the original search results page is subjected to text extraction and fragment segmentation to generate multiple candidate text fragments; Based on the association between the default referential target and the context keywords in the default referential hypothesis, text fragments that co-occur with the association are selected from the multiple candidate text fragments, and the selected text fragments are deduplicated and aggregated to generate a group of accompanying trace fragments corresponding to the default referential hypothesis.

5. The machine learning-based search method according to claim 1, characterized in that, After the accompanying trace fragment group is vectorized and encoded, multiple fragment feature vectors are generated. The multiple fragment feature vectors are input into the machine learning closed loop recognition model. The semantic association strength between fragment feature vectors from different information sources is calculated through the attention mechanism of the machine learning closed loop recognition model. The fragment feature vectors are then clustered according to the semantic association strength to form multiple candidate closed loop clusters.

6. The machine learning-based search method according to claim 5, characterized in that, For each candidate closed loop cluster, extract the implicit target entity that is commonly pointed to by cross-information source fragments in the candidate closed loop cluster, and construct the association path between the implicit target entity and the context keywords in the default referential hypothesis. If the association path forms a closed semantic loop, then the candidate closed loop cluster is determined to be a default referential closed loop. Extract the target entity identifier and the key related words on the associated path from the default referential closed loop, and combine the target entity identifier and the key related words according to the preset query syntax to generate a closed loop search expression.

7. The machine learning-based search method according to claim 1, characterized in that, The closed-loop search expression is decomposed into verification target words and verification related words. The verification target words are derived from the target entity identifier words in the default referencing closed loop, and the verification related words are derived from the key related words in the default referencing closed loop. The verification target word and the verification related word are combined into a verification query, and a verification search is performed in the multiple candidate information sources to obtain the verification search results page returned by each candidate information source. The verification search results page is parsed to extract verification result items that contain the verification target word and also contain the verification related word or its synonym, and a candidate result set is constructed.

8. The machine learning-based search method according to claim 7, characterized in that, Calculate the semantic matching degree between each verification result item in the candidate result set and the default referential loop, filter out verification result items with matching degree lower than a preset threshold, and sort the remaining verification result items in descending order of matching degree to generate a target result sequence consistent with the default referential loop.

9. A machine learning-based search system for implementing the search method as described in any one of claims 1-8, characterized in that, Includes the following modules: Missing Name Target Parsing Module: Performs missing name target parsing on the original search request and generates a default reference hypothesis corresponding to the original search request; Non-target search and trace extraction module: Based on the default reference hypothesis, trigger non-target word search in multiple candidate information sources to extract the accompanying trace fragment group corresponding to the default reference hypothesis; The closed-loop identification and expression generation module inputs the accompanying trace fragment group into the machine learning closed-loop identification model, identifies the default referential closed loops that point to the same implicit target across information sources, and generates a closed-loop search expression based on the default referential closed loops. The closed-loop verification and result output module is used to perform closed-loop verification search on the candidate information source using the closed-loop search expression, and output the target result sequence that is consistent with the default referent closed loop.

10. A computer-readable storage medium, characterized in that, The storage medium stores a computer program that can be loaded by a processor and executed according to any one of claims 1 to 8.

Images