A smart text retrieval method, system, storage medium, and program product

Abstract

This application provides an intelligent text retrieval method, system, storage medium, and program product, relating to the field of intelligent text retrieval. The method includes: segmenting each original document into multiple semantic segments based on semantic similarity; extracting text features from each semantic segment to generate a text segment vector; selecting relevant segments from each text segment vector whose first similarity to the retrieval vector is greater than a preset first threshold; calculating a second similarity between each relevant segment and its adjacent segments; merging adjacent segments with a second similarity greater than a preset second threshold with their corresponding relevant segments to obtain a stitched segment; determining the retrieval granularity to be displayed first based on the retrieval results corresponding to the user's historical retrieval requests; adjusting the display content of the stitched segment according to the retrieval granularity; and returning the text retrieval results. Implementing this method can improve the semantic completeness and relevance of the text retrieval results.

Description

Technical Field

[0001] This application relates to the field of intelligent text retrieval, and more particularly to an intelligent text retrieval method, system, storage medium, and program product. Background Technology

[0002] With the rapid development of information technology, human society has entered an era of data explosion. Whether it's academic papers in scientific research, case files in the legal industry, or technical documents and market reports within enterprises, massive amounts of complex unstructured text data have become the primary carriers of information. Against this backdrop, how to quickly and accurately extract the required information from this vast sea of ​​documents has become a key bottleneck for improving decision-making efficiency and promoting innovative development across all industries. Traditional text retrieval technologies based on keyword matching, due to their inability to deeply understand the inherent semantics of text, are no longer sufficient to meet users' increasingly high standards for information accuracy and completeness.

[0003] In related technologies, each long document is typically divided into multiple independent, standardized text blocks according to preset rules, such as fixed character length, paragraph marks, or chapter structure. Then, the system uses an advanced text vectorization model (such as a language model based on the Transformer architecture) to calculate the semantic information contained within each text block, abstracting and converting it into a high-dimensional mathematical vector. When a user submits a search request in natural language, the system similarly converts this request into a query vector. Finally, by calculating the semantic similarity (e.g., cosine similarity) between this query vector and all text block vectors in the document library in the vector space, the system selects one or more text blocks that are closest to the user's intent and presents these text blocks as the final search results to the user.

[0004] However, the underlying logic of related technologies is to treat the original document as a collection of unrelated text blocks. The retrieval process essentially selects the individual text blocks from this collection that are most directly relevant to the user's query. When a complete argument, a complex operational procedure, or a conclusion with preconditions is precisely divided and distributed across two or more physically consecutive text blocks in the original document, related technologies exhibit their limitations. For example, the system might successfully retrieve the first text block because of its extremely high semantic match with the query, but discard the second text block, which provides supplementary information, limiting conditions, or causal explanations, because its direct relevance to the query is slightly lower. This approach directly undermines the inherent integrity of the retrieval results. Summary of the Invention

[0005] This application provides an intelligent text retrieval method, system, storage medium, and program product to address the problem that multiple related text fragments returned by relevant retrieval systems often have logical inconsistencies or incomplete information after being combined.

[0006] Firstly, this application provides an intelligent text retrieval method, applied to an intelligent text retrieval system, the method comprising:

[0007] Each original document is segmented into multiple semantic segments based on the semantic similarity of its semantic content;

[0008] Extract the text features of each semantic segment to generate a text segment vector corresponding to each semantic segment;

[0009] One or more relevant segments are selected from each text segment vector that have a first similarity greater than a preset first threshold with the retrieval vector, wherein the retrieval vector is converted according to the user's retrieval request;

[0010] Calculate a second similarity between each of the relevant segments and adjacent segments, wherein the adjacent segments and the relevant segments are physically contiguous in the original document;

[0011] The adjacent segments with a similarity greater than a preset second threshold are merged with the corresponding related segments to obtain the stitched segments;

[0012] The search granularity displayed first is determined based on the search results corresponding to the user's historical search requests. The search granularity includes the abstract level, paragraph level, and article level.

[0013] Adjust the displayed content of the suture fragment according to the retrieval granularity, and return the text retrieval results.

[0014] Through the above embodiments, the system subdivides the original document into multiple semantic segments based on semantic similarity, and extracts textual features to generate vectors for each semantic segment, achieving fine-grained, semantic-level modeling of the document content. Based on this, relevant segments are filtered by comparing the similarity between the retrieval vector and the segment vector, and further stitched together by considering the semantic continuity of adjacent segments, ensuring the contextual integrity and semantic coherence of the retrieval results. Finally, the retrieval granularity is dynamically adjusted based on the user's historical behavior, optimizing the adaptability of the results and making the returned content both accurate and comprehensive. This solution improves the semantic integrity and relevance of text retrieval results.

[0015] In some embodiments, the step of segmenting each original document into multiple semantic fragments based on the semantic similarity of its semantic content specifically includes:

[0016] The original document is initially segmented at the sentence level to obtain multiple preliminary sentence units. The preliminary sentence unit is the smallest grammatical unit obtained by dividing the original document based on sentence boundaries.

[0017] Calculate the semantic similarity between each pair of adjacent initial sentence units to obtain the similarity between adjacent sentences;

[0018] The initial sentence units with a similarity greater than a preset threshold between adjacent sentences are merged to obtain multiple semantic segments.

[0019] Through the above embodiments, the system performs preliminary sentence-level segmentation on the original document and merges adjacent sentences into semantic fragments based on their semantic similarity. This approach balances syntactic boundaries and semantic continuity, effectively preventing important semantics from being fragmented into scattered pieces, and improving the accuracy of semantic retrieval and the completeness of the context.

[0020] In some embodiments, after the step of adjusting the display content of the stitched fragment according to the retrieval granularity and returning the text retrieval results, the method further includes:

[0021] In response to a user clicking a preset first control, a document structure graph of the original document corresponding to each of the relevant segments is constructed, and each semantic segment in the original document is used as a node in the document structure graph;

[0022] In each document structure graph, the logical relationships between nodes are identified based on preset logical rules, resulting in one or more logical relationship edges;

[0023] Starting from the node corresponding to each relevant segment, traverse one or more logical relationship edges in the document structure graph corresponding to the starting point to obtain one or more logical paths;

[0024] The semantic fragments corresponding to all nodes on one or more logical paths are aggregated to obtain the stitched fragment;

[0025] The displayed content is regenerated based on the stitched fragment, and the text retrieval results are returned.

[0026] Through the above embodiments, after generating text retrieval results, if the user needs to conduct further in-depth searches, they can regenerate new stitched fragments by clicking the first control on the system interface. This involves further constructing a structural graph of the original document based on selected relevant fragments with high similarity to the retrieval vector, traversing logical relationships to form logical paths, and then aggregating the fragments along those paths to form stitched fragments. This feature, through structured modeling and the mining of logical relationships, not only considers semantic relevance but also combines the logical hierarchy and content connections within the document, which can, to some extent, reduce the omission of semantic fragments caused by relying solely on physical adjacency. Ultimately, this allows the retrieval results to cover the complete logical chain, improving the systematicity and coherence of the information.

[0027] In some embodiments, after the step of adjusting the display content of the stitched fragment according to the retrieval granularity and returning the text retrieval results, the method further includes:

[0028] In response to a user clicking a preset second control, for each relevant segment, candidate segments in other original documents with a semantic similarity greater than a preset third threshold are retrieved;

[0029] Calculate the contextual coherence score for each pair of related segments and candidate segments;

[0030] The relevant segments and candidate segments with context coherence scores greater than a preset coherence threshold are merged to obtain a stitched segment across documents;

[0031] The content to be displayed is regenerated based on the stitched fragments across documents, and the text retrieval results are returned.

[0032] Through the above embodiments, after generating text search results, if the user needs to conduct a broader search, they can regenerate cross-document stitched fragments by clicking the second control on the system interface. That is, during the fragment stitching process, it is not limited to the original document but also introduces candidate fragments from other documents. By calculating the contextual coherence score, multiple related fragments and candidate fragments are merged. Utilizing both semantic similarity and contextual coherence, cross-document semantic splicing is achieved, effectively breaking down document boundary limitations. Ultimately, the search results are improved in terms of information coverage and semantic continuity, avoiding information fragmentation.

[0033] In some embodiments, the step of determining the search granularity to be displayed first based on the search results corresponding to the user's historical search requests specifically includes:

[0034] The average reading dwell time for each granularity of user's historical search requests was statistically analyzed to obtain granularity dwell time data.

[0035] Obtain the similarity between the current search request and historical search requests in terms of topic type and text structure complexity to obtain a request similarity score;

[0036] The historical retrieval granularity corresponding to the historical retrieval request with the highest request similarity score is determined as the retrieval granularity displayed first.

[0037] Through the above embodiments, the system intelligently determines the search granularity to be displayed first by statistically analyzing the dwell time at different search granularities under users' historical search requests and calculating similarity scores based on topic and structural complexity. This dynamic matching based on historical behavior and request characteristics can adapt to users' reading habits and preferences. Ultimately, the granularity of search result display better matches user needs, improving the search experience and content utilization efficiency.

[0038] In some embodiments, the step of transforming the retrieval vector according to the user's retrieval request includes:

[0039] The query text corresponding to the user's search request is segmented into words to obtain the segmentation results;

[0040] By removing stop words from the word segmentation results based on the stop word list, a valid word segmentation set is obtained;

[0041] Based on the effective word segmentation set and all semantic segments in the vector database, the keyword score of each effective word segment is calculated using the TF-IDF algorithm to obtain a word segmentation score set, which is a list of scores corresponding to each effective word segment.

[0042] The word segmentation score set is sorted, and low-scoring words with scores below a preset proportion are removed to obtain the key word segmentation set;

[0043] Generate a retrieval vector corresponding to each valid word in the key word segmentation set.

[0044] Through the above embodiments, the system incorporates steps such as word segmentation, stop word removal, TF-IDF keyword scoring, score sorting, and filtering when converting user search requests into search vectors. This ensures that the final search vector accurately reflects the user's intent. Crucially, score filtering eliminates noisy words while retaining high-weighted words, resulting in a more precise semantic expression in the converted search vector. Ultimately, this improves the matching degree and relevance between the search request and the document fragment.

[0045] In some embodiments, the step of adjusting the displayed content of the stitched fragment according to the retrieval granularity and returning the text retrieval results specifically includes:

[0046] The keywords extracted from the search request are compared with the text content of the suture fragment to obtain the keyword matching degree of each suture fragment. The keyword matching degree is the coverage of the search keywords contained in the suture fragment.

[0047] The display order of different suture fragments is determined according to the keyword matching degree from large to small, and returned as the text search results.

[0048] Through the above embodiments, the system compares search keywords with the content of the stitched fragments, calculates the keyword matching degree, and displays the search results in order of ranking. This not only ensures a high degree of relevance of the results but also improves the intuitiveness and information utilization efficiency of the search results. The introduction of keyword matching degree effectively prevents irrelevant content from being mixed in with the search results. Ultimately, the search results are more targeted and readable, meeting the needs of high-quality information filtering.

[0049] Secondly, this application provides an intelligent text retrieval system, which includes: one or more processors and a memory;

[0050] The memory is coupled to the one or more processors. The memory is used to store computer program code, which includes computer instructions. The one or more processors call the computer instructions to enable the intelligent text retrieval system to implement the intelligent text retrieval method provided in the above embodiments, which will not be described in detail here.

[0051] Thirdly, this application provides a computer-readable storage medium including instructions that, when executed on an intelligent text retrieval system, enable the intelligent text retrieval system to implement an intelligent text retrieval method provided in the above embodiments, which will not be elaborated further here.

[0052] Fourthly, this application provides a computer program product that, when run on an intelligent text retrieval system, enables the intelligent text retrieval system to implement an intelligent text retrieval method provided in the above embodiments, which will not be elaborated here.

[0053] One or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages:

[0054] 1. By segmenting documents based on semantic similarity and combining the semantic continuity of adjacent segments with the logical relationships within the document structure, potentially fragmented and disjointed related segments are dynamically stitched together into semantically complete segments. In particular, the introduction of logical path traversal and cross-document contextual coherence calculation ensures that the final search results not only contain directly related information but also automatically piece together complete arguments or causal chains, thus addressing the technical challenges of logical breaks and information loss after segment combination in traditional retrieval.

[0055] 2. To address the diverse information completeness needs of different users in various search scenarios, the system intelligently determines the optimal display granularity by analyzing the actual reading time and request similarity of users' historical search granularities. In this way, whether users need summaries, paragraphs, or entire articles, they can obtain the best information coverage while ensuring semantic coherence, effectively mitigating the problem of incomplete information caused by overly fine or coarse search results.

[0056] 3. In the process of generating retrieval vectors and ranking results, this solution utilizes multi-dimensional semantic analysis techniques such as word segmentation, TF-IDF weighting, keyword filtering, and coverage matching to ensure that every stitched or combined fragment is closely related to the user's intent. Ranking results by keyword matching further enhances the targeted relevance and usability of the combined content, prevents irrelevant fragments from being mixed in, and ultimately makes the retrieval output not only logically coherent but also complete in content. Attached Figure Description

[0057] Figure 1 This is a flowchart illustrating an intelligent text retrieval method according to an embodiment of this application;

[0058] Figure 2 This is another flowchart illustrating an intelligent text retrieval method in an embodiment of this application;

[0059] Figure 3 This is a schematic diagram of the physical device structure of an intelligent text retrieval system in an embodiment of this application. Detailed Implementation

[0060] The terminology used in the following embodiments of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application. As used in the specification and appended claims of this application, the singular expressions “a,” “an,” “the,” “the,” “the,” and “this” are intended to include the plural expressions as well, unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in this application refers to any or all possible combinations including one or more of the listed items.

[0061] Hereinafter, the terms "first" and "second" are used for descriptive purposes only and should not be construed as implying or suggesting relative importance or implicitly indicating the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature, and in the description of the embodiments of this application, unless otherwise stated, "multiple" means two or more.

[0062] For ease of understanding, the method provided in this implementation is described in process below. Please refer to [link / reference]. Figure 1 This is a flowchart illustrating an intelligent text retrieval method in an embodiment of this application.

[0063] S101. Divide each original document into multiple semantic segments based on the semantic similarity of its semantic content.

[0064] Among them, the original document refers to the initial text file that has not been processed by the system and serves as the data source for retrieval, such as academic papers, case files, and technical documents; semantic content refers to the inherent meaning, ideas, or information expressed by the text fragment; semantic similarity refers to the degree of closeness or correlation between two text fragments in terms of semantic content, which is usually obtained through quantitative calculation; semantic fragment refers to a relatively complete and coherent text unit formed after segmentation based on semantic similarity.

[0065] This step is performed in the initial stage of the intelligent text retrieval system's preprocessing of the original documents, completed before the user initiates a search request. Specifically, the system acquires all the original documents to be retrieved. For each document, instead of relying on surface formatting such as fixed character lengths and paragraph marks for segmentation, it delves into the semantic logic within the document. By calculating the semantic similarity between different text sections in the document, it merges semantically closely related parts and splits semantically less related parts, ultimately obtaining multiple semantic fragments. Each semantic fragment is semantically coherent and can fully express a relatively independent viewpoint, description, or logical unit.

[0066] Optionally, the original document is segmented into sentences to obtain independent sentence units; a pre-trained language model (such as Word2Vec) is used to generate sentence vectors for each sentence; the cosine similarity of the sentence vectors of adjacent sentences is calculated as semantic similarity; a semantic similarity threshold is set, and consecutive adjacent sentences with similarity greater than the threshold are merged into a semantic segment until the entire document is traversed.

[0067] Optionally, the original document is initially divided into paragraphs to obtain initial paragraph units; the core keywords and semantic topics of each paragraph are extracted; the matching degree of the core topics of adjacent paragraphs (such as the Jaccard coefficient) is calculated as semantic similarity; based on a preset threshold, adjacent paragraphs with high topic matching degree are merged into semantic segments, and paragraphs with large topic differences are used as the starting point of new segments.

[0068] It is understandable that other methods can be used to achieve semantic segmentation, such as combining semantic role labeling to identify core semantic components before segmentation, which is not limited here.

[0069] S102. Extract the text features of each semantic segment and generate the text segment vector corresponding to each semantic segment.

[0070] Among them, text features refer to attributes or elements that can characterize the core information of semantic segments, including keywords, parts of speech, syntactic structure, semantic topics, etc.; text segment vectors refer to high-dimensional vectors obtained by mathematically transforming the text features of semantic segments, which are used to quantify the semantic information of segments, so that semantic similarity can be calculated through vector operations.

[0071] Specifically, for each semantic segment obtained in step S101, the system uses a feature extraction algorithm to mine its deep text features. These features include not only surface keyword frequencies but also abstract information such as semantic associations and contextual dependencies. Subsequently, a vector transformation model (such as a Transformer-based language model) is used to map these features into high-dimensional text segment vectors. The position of each vector in mathematical space reflects the semantic connotation of its corresponding semantic segment, and the distance or angle between vectors can be used to measure semantic similarity.

[0072] Optionally, the semantic segments are segmented and stop words are removed to retain effective vocabulary; the processed text is encoded using a BERT pre-trained model to obtain the contextual features of the segments; and the [CLS] token vector of the last layer of the model is extracted as the text segment vector of the semantic segment.

[0073] Optionally, the TF-IDF weight of each word in the semantic segment is calculated, and high-weight keywords are selected; word vectors of each keyword are generated based on the Word2Vec model; and the word vectors of the keywords are weighted and averaged using the TF-IDF weights to obtain the text segment vector of the semantic segment.

[0074] It is understandable that other methods can be used to generate text fragment vectors, such as directly generating sentence-level vectors using Sentence-BERT, which is not limited here.

[0075] S103. Select one or more related segments from each text segment vector whose first similarity to the retrieval vector is greater than a preset first threshold.

[0076] The retrieval vector is a vector obtained by transforming a user's retrieval request using the same method as text fragment vectors, and is used to quantitatively represent the user's retrieval intent.

[0077] Specifically, the system first receives a user's retrieval request (such as a natural language query) and converts it into a retrieval vector according to the same logic as that for generating text fragment vectors. Subsequently, it traverses all the text fragment vectors generated in S102 and calculates the first similarity (common calculation methods such as cosine similarity, Euclidean distance, etc.) between each text fragment vector and the retrieval vector. It compares the calculated first similarity with a preset first threshold, and filters out the semantic fragments whose similarity exceeds the threshold. These fragments are the relevant fragments directly related to the user's retrieval request.

[0078] Optionally, perform word segmentation and stop word removal on the user's retrieval request to generate a query text; use the same BERT model as that for text fragment vectors to generate a retrieval vector; calculate the cosine similarity between each text fragment vector and the retrieval vector as the first similarity; set a preset first threshold (such as 0.7), and mark the fragments with a similarity greater than 0.7 as relevant fragments.

[0079] Optionally, calculate the weights of the keywords in the retrieval request based on the TF-IDF algorithm to generate a retrieval vector; calculate the weighted Euclidean distance between the text fragment vector and the retrieval vector, and the smaller the distance, the higher the first similarity; set a distance threshold (such as 0.3), and mark the fragments with a distance less than 0.3 as relevant fragments; preliminarily sort the relevant fragments in descending order of similarity.

[0080] It can be understood that other methods can also be used to implement the screening of relevant fragments. For example, a comprehensive judgment can be made by combining keyword exact matching and semantic similarity, which is not limited here.

[0081] In addition, the system can also receive the query text input by the user, split the continuous text into discrete words or subwords through an adapted word segmentation tool (such as Jieba for Chinese and WordPiece for English) to obtain the word segmentation result, providing a basic semantic unit for subsequent processing; then, based on a general and domain-specific stop word list, filter out the meaningless words (such as "de", "zai") in the word segmentation result, and retain the effective word segmentation set that can represent the core intention, reducing noise interference; subsequently, using the TF-IDF algorithm, combine the term frequency (TF) of the effective word segmentation in the query and the inverse document frequency (IDF) in all semantic fragments in the vector database to calculate the keyword score of each effective word segmentation, generating a set of word segmentation scores to quantify the importance of words; then sort the set of word segmentation scores in descending order of scores, remove the low-score words according to a preset ratio (such as the top 60%), and screen out the core key word segmentation set to focus on the core retrieval intention; finally, call a model consistent with the generation of semantic fragment vectors (such as GloVe, BERT) to generate word vectors for the key word segmentation, and generate the final retrieval vector through weighted fusion (according to the keyword score weights) or context encoding, realizing the accurate quantification of the retrieval intention and laying a foundation for subsequent similarity matching.

[0082] S104. Calculate the second similarity between each relevant segment and its adjacent segments.

[0083] Specifically, for each relevant segment selected in step S103, the system traces back to its corresponding original document and locates the semantic segments (i.e., adjacent segments) that are physically adjacent to the relevant segment. Then, using the same logic as calculating semantic similarity (such as vector cosine similarity, topic matching, etc.), the system calculates the degree of semantic association between the relevant segment and each adjacent segment, i.e., the second similarity. The level of the second similarity reflects the semantic coherence between the relevant segment and its adjacent segments.

[0084] S105. Merge adjacent segments with a second similarity greater than a preset second threshold with their corresponding related segments to obtain stitched segments.

[0085] Specifically, for each relevant segment, the system compares its second similarity with that of adjacent segments with a preset second threshold. If the second similarity exceeds the threshold, it indicates that the two are highly related semantically and logically coherent, and the system merges the adjacent segment with the relevant segment; if it does not exceed the threshold, it is not merged. Through this process, previously scattered relevant segments are "stitched" into semantically more complete segments, preserving core related content while supplementing necessary contextual information, thus avoiding information loss caused by segmentation.

[0086] Optionally, each relevant segment and its corresponding preceding and following adjacent segments are traversed; adjacent segments with a second similarity greater than a preset second threshold (e.g., 0.65) are extracted; the relevant segment is sequentially spliced ​​with its preceding and following adjacent segments that meet the conditions to form a stitched segment containing context; redundant information (e.g., repeated expressions) is removed from the merged stitched segment to retain the core semantics.

[0087] Optionally, for each relevant segment, retrieve all adjacent segments in the same document that are physically continuous with the segment; sort the adjacent segments from high to low according to the second similarity; sequentially merge the adjacent segments that are ranked first and have a second similarity greater than a preset second threshold with the relevant segments until the second similarity of the newly added adjacent segments is lower than the threshold; generate the merged stitched segment and mark its source location.

[0088] It is understandable that other methods can be used to merge fragments, such as combining semantic role labeling to identify core semantic components before merging, which is not limited here.

[0089] S106. Determine the granularity of the search results to be displayed first based on the search results corresponding to the user's historical search requests.

[0090] Among them, historical search requests refer to the search query text and related parameters initiated by users in the past; search results refer to the system's returned content corresponding to historical search requests; search granularity refers to the level of detail displayed in the search results, used to adapt to users' needs for information completeness; summary level refers to the display granularity in the form of a brief summary and extraction of core viewpoints, with concise and refined content; paragraph level refers to the granularity of display based on paragraphs, preserving the paragraph structure and key details of the text; and chapter level refers to the granularity of display based on complete chapters, including the complete logical chain and all details of the text.

[0091] Specifically, the system retrieves the user's historical search data, including the topic and text structure of historical search requests, as well as the user's feedback on different granularities (such as reading time) in the corresponding search results. By analyzing historical data, the system calculates the similarity between the current search request and historical search requests in dimensions such as topic type and text complexity, and finds the most similar historical request. Subsequently, the granularity of the search preferred by the user in that historical request (such as the granularity with the longest reading time) is used as the granularity displayed first in the current search, ensuring that the display format matches the user's habits.

[0092] Optionally, extract all historical search requests and their corresponding search granularities (summary level / paragraph level / article level) and user reading dwell time from the user's historical data; calculate the average dwell time for each granularity in historical requests, and select the granularity with the longest dwell time as the "preference granularity" of the historical request; calculate the similarity between the current search request and each historical request in terms of topic keywords, text length, and structural complexity to obtain a request similarity score; and determine the "preference granularity" of the historical request with the highest similarity score as the search granularity displayed first.

[0093] S107. Adjust the displayed content of the stitched fragments according to the retrieval granularity and return the text retrieval results.

[0094] Specifically, the system makes targeted adjustments to the stitched fragments based on the determined retrieval granularity: at the abstract level, it extracts the core viewpoints and key data of the stitched fragment for concise summarization; at the paragraph level, it retains the original paragraph structure of the fragment, highlighting keywords and key sentences; and at the article level, it presents the complete content of the stitched fragment and supplements necessary contextual explanations. Simultaneously, the system calculates the matching degree between each stitched fragment and the search query keywords, sorts them from highest to lowest matching degree, and finally returns the adjusted content as the text search result to the user.

[0095] Optionally, the content of the stitched fragments can be trimmed according to the retrieval granularity: at the summary level, retain the core sentences (no more than 3 sentences) directly related to the search keywords; at the paragraph level, retain the complete paragraph structure and highlight the keywords; at the chapter level, retain all content and add logical connectors; extract the keywords from the search request, calculate the keyword coverage rate (including the number of keywords / total number of keywords) of each stitched fragment as the matching degree; sort the stitched fragments from high to low matching degree, and display the top 5 results first; organize the sorted results according to the granularity format, add source identifiers, and return them to the user.

[0096] Optionally, corresponding content generation modules can be invoked for different granularities: the summary-level module generates a structured summary of the fragment (including background, core conclusions, and key evidence); the paragraph-level module preserves the original paragraph layout and marks the paragraph topic sentences; the chapter-level module generates a table of contents and a core viewpoint index for the fragment; calculates the keyword matching degree and semantic similarity comprehensive score for each stitched fragment; sorts the results by comprehensive score and filters the top 10 results; adds a "switch granularity" option to the results, allowing users to dynamically adjust the display format, and finally returns the sorted results.

[0097] It is understandable that other methods can be used to adjust the displayed content and return results, such as optimizing the layout style based on the user's browsing history; this is not limited here.

[0098] Furthermore, after the system returns the retrieval results based on the stitched segments determined by the second similarity between each relevant segment and adjacent segments in steps S104 and S105, if the user is not satisfied with the current results, they can perform further retrieval by clicking the selection controls (including the preset first control and the preset second control) on the system interface: The first control is used to further mine the content of each original document in depth, and to mine the relevant segments in depth by constructing a document structure diagram and the logical relationships between different nodes in the document structure diagram from the original documents corresponding to each relevant segment, so that the returned retrieval results are more comprehensive and complete; The second control is used to further mine the breadth of each relevant segment, and to splice the relevant segments of different documents by calculating the semantic similarity and coherence score with relevant segments of other documents, thereby expanding the retrieval results and improving the "breadth" of the returned retrieval results.

[0099] For details on the execution process of clicking the preset first control, please refer to steps S108 to S111.

[0100] S108. In response to the user clicking the preset first control, construct the document structure diagram of the original document corresponding to each relevant segment.

[0101] Specifically, for each relevant segment, the system locates its original document and then extracts all semantic segments (including relevant segments and other irrelevant semantic segments) from that original document as nodes in the document structure graph. Node attributes include the segment's text content, its physical location in the original document (such as start / end page numbers, paragraph numbers), and core topic tags. Simultaneously, the system initially defines the basic connections between nodes (such as physical adjacency relationships), forming a preliminary document structure graph framework.

[0102] Optionally, for the original document corresponding to the relevant fragment, extract all semantic fragments and assign a unique node ID to each fragment; add attribute information to each node: text content, physical location index, and core keyword set; connect nodes with consecutive positions using "adjacent" edges based on the physical order of the fragments in the original document; and save the document structure graph containing node attributes and adjacent edges to the graph database.

[0103] Optionally, a graph building tool (such as Neo4j) can be used to create a graph space based on the original document; all semantic fragments of the document can be converted into nodes in the graph, and the node ID can be bound to the unique identifier of the fragment in the document; using the semantic similarity calculation results, "semantic association" edges can be added to non-adjacent fragments with semantic similarity higher than a basic threshold (such as 0.3); and a document structure graph containing nodes, adjacent edges, and semantic association edges can be generated and visualized.

[0104] S109. In each document structure diagram, the logical relationships between nodes are identified based on preset logical rules to obtain one or more logical relationship edges.

[0105] Specifically, for each document structure graph, the system calls upon a pre-defined logical rule library to analyze the text content of all nodes (semantic fragments) in the graph. Logical rules can include: rules based on logical connectors (e.g., "because…therefore…" corresponding to causal relationships), rules based on syntactic structure (e.g., "if…then…" corresponding to conditional relationships), and rules based on topic progression (e.g., "first…second…finally…" corresponding to progressive relationships). By matching these rules, the system identifies the specific logical relationship types between nodes and adds logical relationship edges with relationship type labels (e.g., "causal edge" or "progressive edge") between the corresponding nodes. These edges accurately reflect the logical coherence between fragments, enhancing the semantic expressiveness of the document structure graph.

[0106] S110. Starting from the node corresponding to each relevant segment, traverse one or more logical relationship edges in the document structure graph corresponding to the starting point to obtain one or more logical paths.

[0107] Specifically, for each relevant segment's corresponding node (starting node), the system initiates a traversal process within its associated document structure graph. The traversal methods include depth-first traversal (prioritizing exploration of single logical branches) and breadth-first traversal (prioritizing coverage of multiple logical branches). The system adjusts the traversal priority based on the importance of logical relationship types (e.g., causal relationships take precedence over parallel relationships). During traversal, the system sequentially visits associated nodes along logical relationship edges, recording the node ID and the logical relationship type of the edge, until it encounters a node with no subsequent logical relationship edges or a cyclic node. Ultimately, this forms one or more logical paths extending from the starting node, each path containing a continuous, logically related sequence of nodes.

[0108] Optionally, determine the starting node corresponding to the relevant segment, initialize the path list and add the starting node to the current path; identify all logical relationship edges associated with the starting node and sort them according to the importance of the relationship (e.g., causal > progressive > parallel); visit the next node along the first sorted logical relationship edge and add that node to the current path; repeat the sorting steps until there are no untraversed logical relationship edges for the node, save the current path and backtrack to the nearest branch point, continue to traverse other branches, and finally obtain multiple logical paths.

[0109] Optionally, set weights for logical relationship edges in the document structure graph (e.g., causal relationship weight is 0.8, parallel relationship weight is 0.5); use Dijkstra's algorithm to calculate the shortest weighted path to other nodes with the starting node as the center (the lower the weight, the higher the path priority); filter paths with weighted path length less than a preset threshold (e.g., weight sum < 2.0) as valid logical paths; sort by path weight sum from smallest to largest, and retain the first N (e.g., the first 3) logical paths.

[0110] S111. Gather the semantic fragments corresponding to all nodes on one or more logical paths to obtain the stitched fragment.

[0111] Specifically, for each logical path, the system extracts the semantic fragment text content corresponding to all nodes on the path. The fragments are sorted according to the physical order of the nodes in the original document (or the traversal order within the logical path) to ensure the coherence of the text flow. After sorting, the system optimizes the connections between fragments, such as removing repetitive expressions and adding logical transition words (e.g., "therefore" or "in addition") to enhance coherence. For multiple logical paths, the system can select the optimal path's corresponding fragment aggregation based on the path's completeness (e.g., the number of nodes included) and relevance to the retrieval request (e.g., keyword coverage), or merge the core fragments of multiple paths to ultimately form a stitched-together fragment.

[0112] Finally, the system regenerates the display content based on the final suture fragment and returns the search results. The process of regenerating the display content based on the final suture fragment is basically the same as the data processing process in step S107, and will not be described again here.

[0113] In addition, the execution process of clicking the preset first control is as follows:

[0114] The system, for each relevant fragment selected from the original document, breaks through the limitation of a single document and expands the search scope to all other original documents. By calculating the semantic similarity between the relevant fragment and semantic fragments in other documents, candidate fragments with similarity greater than a preset third threshold are selected, and potential supplementary information is mined from the global document library. Next, for each pair of relevant fragments and candidate fragments, analysis is performed from multiple dimensions such as topic consistency, logical relationship, and expression style. The results of each dimension are quantified and weighted and integrated to obtain a quantitative score that measures the contextual coherence of the two, ensuring that the selected candidate fragments are not only semantically similar but also logically coherent. Finally, the system compares the contextual coherence score with a preset coherence threshold, splices fragment pairs with scores exceeding the threshold in logical order, adds transition words and removes duplicate expressions when necessary, and finally forms a cross-document stitched fragment that integrates relevant information from multiple documents, is semantically coherent, and has complete information.

[0115] Finally, the system regenerates the display content based on the cross-document stitched fragments and returns the search results. The process of regenerating the display content based on the cross-document stitched fragments is basically the same as the data processing process in step S107, and will not be described in detail here.

[0116] The following provides a more detailed description of the process of the method provided in this implementation. Please refer to [link / reference]. Figure 2 This is another flowchart illustrating an intelligent text retrieval method in an embodiment of this application.

[0117] S201. Perform preliminary sentence-level segmentation on the original document to obtain multiple preliminary sentence units.

[0118] This step is performed in the initial stage of semantic segmentation of the original document by the intelligent text retrieval system. Specifically, the system acquires the original document to be processed and uses natural language processing technology to identify sentence boundaries in the document (such as punctuation marks like periods, question marks, and exclamation marks, or syntactic markers of specific languages), splitting the continuous text stream into independent preliminary sentence units according to sentence boundaries. Each preliminary sentence unit retains complete grammatical structure and semantic information to ensure the accuracy of subsequent semantic similarity calculations.

[0119] Optionally, the text content of the original document is read, and redundant spaces, line breaks, and other formatting symbols are removed. Sentence delimiters are located based on preset sentence boundary rules (such as recognizing Chinese "。?!" and English ".", "?", "!"). The text is divided into independent sentence strings by the delimiters, with each string serving as a preliminary sentence unit. The segmented sentence units are then validated, and excessively short fragments caused by punctuation errors (such as punctuation marks after abbreviations) are merged.

[0120] Optionally, the sentence segmentation function in a natural language processing tool (such as NLTK or spaCy) can be called; the original document text can be preprocessed to mark punctuation marks and syntactic pauses; the complete sentences can be identified using the tool's built-in sentence boundary detection model to generate a preliminary list of sentence units; the segmentation results can be manually or automatically sampled to correct missegmented special sentence structures (such as sentences within quotation marks).

[0121] S202. Calculate the semantic similarity between each pair of adjacent preliminary sentence units to obtain the adjacent sentence similarity.

[0122] Specifically, the system iterates through all initial sentence units, selecting each pair of adjacent sentence units sequentially according to the order in the original document (e.g., sentence 1 and sentence 2, sentence 2 and sentence 3, etc.). Using semantic similarity calculation algorithms (such as vector cosine similarity, topic matching, etc.), it analyzes the degree of association between adjacent sentences in terms of semantic content, core theme, and logical relationship, obtaining corresponding similarity scores. Higher scores indicate stronger semantic coherence between adjacent sentences, making them more suitable for merging into a single semantic segment.

[0123] Optionally, extract the core keywords of each preliminary sentence unit (using TF-IDF or TextRank algorithms); calculate the Jaccard coefficient (number of intersection keywords / number of union keywords) of the keyword sets of adjacent sentences as the basic similarity; adjust the similarity score by combining the results of the sentence's syntactic dependency analysis (such as subject-verb-object association); and generate the final similarity score between adjacent sentences.

[0124] It is understandable that other methods can be used to calculate similarity, such as calculating similarity after generating sentence embeddings using the BERT model; this is not a limitation here.

[0125] S203. Merge adjacent sentences to obtain multiple semantic segments, and select one or more relevant segments from them based on the retrieval vector.

[0126] Specifically, the system compares the similarity score of each pair of adjacent initial sentence units with a preset threshold. If the similarity score is greater than the threshold, it indicates that the two sentences are semantically closely related and logically coherent, and the system merges the two sentence units. If the similarity score is less than or equal to the threshold, they are not merged, and the threshold is used as the boundary to define the starting point of a new semantic segment. Through sentence-by-sentence traversal and merging, the originally independent sentence units are integrated into multiple semantic segments. Each segment contains continuous and semantically coherent sentences, which can fully express a relatively independent viewpoint, description, or logical unit.

[0127] Next, the system executes steps S102 and S103 based on the integrated semantic fragments, and selects one or more relevant fragments from them according to the retrieval vector, which will not be described in detail here.

[0128] Optionally, initialize the semantic fragment list, taking the first preliminary sentence unit as the starting sentence of the first fragment; sequentially traverse adjacent sentence pairs, if the similarity is greater than a preset threshold (e.g., 0.6), then merge the next sentence unit into the current fragment; if the similarity is less than or equal to the threshold, then end the current fragment and take the next sentence unit as the starting sentence of the new fragment; after traversal, generate a semantic fragment list containing all merged fragments, and mark the number of sentences in each fragment.

[0129] Optionally, the similarity scores of all adjacent sentences are sorted, and sentence pairs with similarity scores greater than a preset threshold are selected; a sentence association graph is constructed based on the selection results, and adjacent sentence pairs with the same similarity score are connected by edges; connected components in the association graph are identified, and continuous sentence units contained in each connected component are merged into a semantic segment; the semantic integrity of the merged segment is checked, and segments that are too long or have a topic deviation are split.

[0130] S204. Construct a stitched segment based on the second similarity between each relevant segment and its adjacent segments.

[0131] S205. Calculate the average reading dwell time for each granularity of the user's historical search requests to obtain granularity dwell time data.

[0132] Specifically, the system extracts all historical search request records from user behavior logs, including the displayed content at different granularities (summary level, paragraph level, and document level) in the search results for each request, as well as user reading behavior data (such as page dwell time and scroll depth). For each search granularity, the system filters out abnormal dwell times (such as excessively short accidental touch times) and calculates the average dwell time for that granularity. For example, if a user spends an average of 5 minutes on paragraph-level results for a historical request, but only 1 minute on summary-level results, it indicates that the user prefers paragraph-level granularity. Finally, granular dwell time data containing the correspondence between "search granularity - average dwell time" is generated.

[0133] S206. Obtain the similarity between the current search request and historical search requests in terms of topic type and text structure complexity, and get the request similarity score.

[0134] Specifically, the system performs topic type labeling (e.g., through keyword matching or topic model classification) and text structure complexity analysis (e.g., calculating average sentence length and terminology ratio) on the current search request. Then, it iterates through historical search requests, extracting the topic type and structure complexity features of each historical request. By calculating the matching degree between the current request and historical requests in terms of topic type (e.g., 1 point for the same topic, 0.6 points for related topics) and the Euclidean distance in terms of structure complexity (the smaller the distance, the higher the similarity), the results from the two dimensions are weighted and fused (e.g., topic weight 0.7, structure weight 0.3) to obtain the final request similarity score. The higher the score, the more similar the search scenarios of the two requests.

[0135] Optionally, extract topic keywords for the current search request and historical search requests respectively (e.g., using the TF-IDF algorithm); calculate the Jaccard similarity of the topic keyword set as the topic matching degree; analyze the text structure complexity, calculate the average sentence length difference and term density difference between the current and historical requests, and obtain the structural similarity after normalization; and obtain the request similarity score by weighted summation according to topic matching degree (weight 0.6) and structural similarity (weight 0.4).

[0136] S207. The historical retrieval granularity corresponding to the historical retrieval request with the highest request similarity score is determined as the retrieval granularity displayed first.

[0137] Specifically, the system sorts all historical search requests from highest to lowest based on their similarity scores, and selects the historical search request with the highest score (i.e., the historical sample most similar to the current request). Then, it extracts the historical search granularity corresponding to that historical request (i.e., the granularity where the user spent the longest time under that historical request, as calculated in step S205), and determines it as the granularity that the current search will display first. For example, if the user spent the longest time at the paragraph level among the historical requests most similar to the current request, then the current search will prioritize displaying paragraph-level content.

[0138] S208. Compare the keywords extracted from the retrieval request with the text content of the suture fragments to obtain the keyword matching degree of each suture fragment.

[0139] Specifically, the system extracts keywords from the current search request (e.g., filtering high-weight words using TF-IDF or TextRank algorithms) to form a keyword set. Then, it iterates through the text content of each stitched segment, counting the number of keywords contained in the segment (including exact matches and synonym matches). The number of matched keywords is divided by the total number of keywords in the search request to obtain the keyword match score of the stitched segment (ranging from 0 to 1). For example, if the search keyword is "artificial intelligence + search method," and a stitched segment contains two keywords, the match score is 1; if it only contains "artificial intelligence," the match score is 0.5.

[0140] Optionally, extract N core keywords (e.g., N=5) from the search request to form a keyword list; perform word segmentation and stop word removal on each stitched segment to obtain a segment vocabulary set; count the number M words in the segment vocabulary set that completely match the keyword list; calculate the matching degree = M / N, and retain two decimal places as the final result.

[0141] Optionally, the search request keywords are expanded using synonyms (e.g., through WordNet or domain thesaurus) to form an expanded keyword set; the stitched fragment text is traversed, and matching keywords are identified through string matching and semantic similarity calculation (e.g., word vector cosine similarity > 0.8); the number of matching expanded keywords M is counted, and duplicate matches are removed; the matching degree is calculated as M / the number of original keywords, where the number of original keywords is the total number of keywords before expansion.

[0142] S209. Determine the display order of different suture fragments according to the keyword matching degree from large to small, and return them as text search results.

[0143] Specifically, the system sorts all the stitched fragments from highest to lowest keyword matching score. Fragments with the same matching score are then sorted by their physical location in the original document (or by semantic similarity). Next, based on the retrieval granularity determined in step S207, the sorted fragments are adjusted (e.g., extracting core sentences at the summary level and preserving the original structure at the paragraph level). Finally, the adjusted fragments are displayed in sorted order on the results page, along with auxiliary information such as source document identifiers and keyword highlighting, forming a complete text search result returned to the user.

[0144] Optionally, all stitched fragments are sorted in descending order of keyword matching degree to generate a preliminary sorted list; fragments with the same matching degree are sorted a second time according to their corresponding logical path length (the longer the better); based on the first search granularity (such as the abstract level), corresponding display content is generated for each fragment; the fragments are displayed in sorted order on the results page, with matching degree scores and source documents labeled, and returned to the user.

[0145] The intelligent text retrieval system of this invention is applied to electronic devices. Figure 3 A schematic diagram of the architecture of an electronic device suitable for implementing embodiments of the present invention is shown.

[0146] It should be noted that, Figure 3 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of use of the embodiments of the present invention.

[0147] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by instructions (computer programs), or by instructions (computer programs) controlling related hardware. These instructions can be stored in a computer-readable storage medium and loaded and executed by a processor. The electronic device of this embodiment includes a storage medium and a processor, wherein the storage medium stores multiple instructions that can be loaded by the processor to execute any step of the method provided in the embodiments of the present invention.

[0148] Specifically, the storage medium and the processor are electrically connected directly or indirectly to enable data transmission or interaction. For example, these components can be electrically connected to each other via one or more signal lines. The storage medium stores computer-executable instructions that implement data access control methods, including at least one software functional module that can be stored in the storage medium in the form of software or firmware. The processor executes various functional applications and data processing by running the software program and module stored in the storage medium. The storage medium can be, but is not limited to, Random Access Memory (RAM), Read-Only Memory (ROM), Programmable Read-Only Memory (PROM), Erasable Programmable Read-Only Memory (EPROM), Electrically Erasable Programmable Read-Only Memory (EEPROM), etc. The storage medium stores the program, and the processor executes the program after receiving the execution instructions.

[0149] Furthermore, the software programs and modules within the aforementioned storage medium may also include an operating system, which may include various software components and / or drivers for managing system tasks (such as memory management, storage device control, power management, etc.) and can communicate with various hardware or software components to provide an operating environment for other software components. The processor can be an integrated circuit chip with signal processing capabilities. The aforementioned processor can be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc., which can implement or execute the methods, steps, and logic flowcharts disclosed in this embodiment. The general-purpose processor can be a microprocessor or any conventional processor.

[0150] Since the instructions stored in the storage medium can execute the steps in any of the methods provided in the embodiments of the present invention, the beneficial effects of any of the methods provided in the embodiments of the present invention can be achieved, as detailed in the preceding embodiments, and will not be repeated here.

[0151] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. An intelligent text retrieval method, applied to an intelligent text retrieval system, characterized in that, The method includes: Each original document is segmented into multiple semantic segments based on the semantic similarity of its semantic content; Extract the text features of each semantic segment to generate a text segment vector corresponding to each semantic segment; One or more relevant segments are selected from each text segment vector that have a first similarity greater than a preset first threshold with the retrieval vector, wherein the retrieval vector is converted according to the user's retrieval request; Calculate a second similarity between each of the relevant segments and adjacent segments, wherein the adjacent segments and the relevant segments are physically contiguous in the original document; The adjacent segments with a similarity greater than a preset second threshold are merged with the corresponding related segments to obtain the stitched segments; The search granularity displayed first is determined based on the search results corresponding to the user's historical search requests. The search granularity includes the abstract level, paragraph level, and article level. Adjust the displayed content of the suture fragment according to the retrieval granularity, and return the text retrieval results.

2. The method according to claim 1, characterized in that, The step of segmenting each original document into multiple semantic fragments based on the semantic similarity of its semantic content specifically includes: The original document is initially segmented at the sentence level to obtain multiple preliminary sentence units. The preliminary sentence unit is the smallest grammatical unit obtained by dividing the original document based on sentence boundaries. Calculate the semantic similarity between each pair of adjacent initial sentence units to obtain the similarity between adjacent sentences; The initial sentence units with a similarity greater than a preset threshold between adjacent sentences are merged to obtain multiple semantic segments.

3. The method according to claim 1, characterized in that, After the step of adjusting the display content of the stitched fragment according to the retrieval granularity and returning the text retrieval results, the method further includes: In response to a user clicking a preset first control, a document structure graph of the original document corresponding to each of the relevant segments is constructed, and each semantic segment in the original document is used as a node in the document structure graph; In each document structure graph, the logical relationships between nodes are identified based on preset logical rules, resulting in one or more logical relationship edges; Starting from the node corresponding to each relevant segment, traverse one or more logical relationship edges in the document structure graph corresponding to the starting point to obtain one or more logical paths; The semantic fragments corresponding to all nodes on one or more logical paths are aggregated to obtain the stitched fragment; The displayed content is regenerated based on the stitched fragment, and the text retrieval results are returned.

4. The method according to claim 1, characterized in that, After the step of adjusting the display content of the stitched fragment according to the retrieval granularity and returning the text retrieval results, the method further includes: In response to a user clicking a preset second control, for each relevant segment, candidate segments in other original documents with a semantic similarity greater than a preset third threshold are retrieved; Calculate the contextual coherence score for each pair of related segments and candidate segments; The relevant segments and candidate segments with context coherence scores greater than a preset coherence threshold are merged to obtain a stitched segment across documents; The content to be displayed is regenerated based on the stitched fragments across documents, and the text retrieval results are returned.

5. The method according to claim 1, characterized in that, The step of determining the search granularity to be displayed first based on the search results corresponding to the user's historical search requests specifically includes: The average reading dwell time for each granularity of user's historical search requests was statistically analyzed to obtain granularity dwell time data. Obtain the similarity between the current search request and historical search requests in terms of topic type and text structure complexity to get the request similarity score; The historical retrieval granularity corresponding to the historical retrieval request with the highest request similarity score is determined as the retrieval granularity displayed first.

6. The method according to claim 1, characterized in that, The steps for transforming the retrieval vector based on the user's retrieval request include: The query text corresponding to the user's search request is segmented into words to obtain the segmentation results; By removing stop words from the word segmentation results based on the stop word list, an effective word segmentation set is obtained; Based on the effective word segmentation set and all semantic segments in the vector database, the keyword score of each effective word segment is calculated using the TF-IDF algorithm to obtain a word segmentation score set, which is a list of scores corresponding to each effective word segment. The word segmentation score set is sorted, and low-scoring words with scores below a preset proportion are removed to obtain the key word segmentation set; Generate a retrieval vector corresponding to each valid word in the key word segmentation set.

7. The method according to claim 1, characterized in that, The step of adjusting the displayed content of the stitched fragment according to the retrieval granularity and returning the text retrieval results specifically includes: The keywords extracted from the search request are compared with the text content of the suture fragment to obtain the keyword matching degree of each suture fragment. The keyword matching degree is the coverage of the search keywords contained in the suture fragment. The display order of different suture fragments is determined according to the keyword matching degree from large to small, and returned as the text search results.

8. An intelligent text retrieval system, characterized in that, The intelligent text retrieval system includes: one or more processors and a memory; The memory is coupled to the one or more processors, the memory being used to store computer program code, the computer program code including computer instructions, the one or more processors invoking the computer instructions to cause the intelligent text retrieval system to perform the method as described in any one of claims 1-7.

9. A computer-readable storage medium comprising instructions, characterized in that, When the instruction is executed on the intelligent text retrieval system, it causes the intelligent text retrieval system to perform the method as described in any one of claims 1-7.

10. A computer program product, characterized in that, When the computer program product is run on the intelligent text retrieval system, the intelligent text retrieval system performs the method as described in any one of claims 1-7.

Images