A multi-dimensional topic discovery method, system, device, medium and product based on prior information and boundary perception

Abstract

The application discloses a multi-dimensional topic discovery method, system, device, medium and product based on prior information and boundary perception, relates to the technical field of natural language processing, and comprises the following steps: acquiring a target text corpus set and a corresponding prior topic information set; performing text segmentation on the corpus to obtain a text unit set; constructing a positive topic prototype and a negative boundary prototype of each prior topic information; determining a topic support level of each prior topic information based on the text unit set, and determining effective prior topic information; performing soft evidence text paragraph retrieval on the effective prior topic information to obtain a candidate text evidence paragraph set; and screening the candidate text evidence paragraph to obtain a topic discovery result. The application solves the problems that traditional topic modeling cannot correspond to a pre-attention dimension, a seed guide method lacks evidence judgment and boundary control, and is prone to generate pseudo topics, and improves the accuracy, evidence reliability and result interpretability of topic discovery.

Description

Technical Field

[0001] This application relates to the field of natural language processing technology, and in particular to a method, system, device, medium and product for multi-dimensional topic discovery based on prior information and boundary awareness. Background Technology

[0002] With the rapid development of the internet and big data technologies, the amount of text data accumulated across various industries has exploded. How to efficiently and accurately identify thematic information that matches the researchers' pre-existing focus from massive amounts of text has become an important research direction in the field of natural language processing.

[0003] Traditional topic modeling methods typically automatically identify high-frequency co-occurring word clusters as topics from corpora. However, the topics discovered by these unsupervised methods are often based on statistical co-occurrence relationships and may not necessarily conform to the conceptual dimensions that researchers are interested in. For example, the model may discover generalized topics such as "system, platform, management, service," but cannot accurately correspond to specific application scenarios such as "R&D design," "production and manufacturing," "supply chain management," and "operation and maintenance services." Another related technology is a topic discovery method based on seed words, where the user provides several seed words, and the system expands related words around the seed words to form topics. However, these methods generally assume that each seed word can form a valid topic, lacking a mechanism to determine whether the topic actually exists in the corpus. When a seed word lacks sufficient textual evidence in the corpus, this type of method will still barely generate a set of seemingly related topic words, leading to the generation of "pseudo-topics." Furthermore, this type of method relies solely on positive seed words for semantic expansion, easily absorbing generalized words and confusing words from adjacent topics, and lacks effective control over topic boundaries. Furthermore, the output of such methods is usually a list of keywords, which cannot be traced back to specific original text evidence paragraphs, resulting in insufficient interpretability of the keywords and making it difficult for users to verify their reliability.

[0004] Therefore, there is an urgent need for a multidimensional topic discovery method based on prior information and boundary awareness to address the problems that traditional topic modeling cannot accurately correspond to the conceptual dimensions that researchers are concerned about in advance, that seed word-guided methods lack evidence judgment mechanisms and are prone to generating pseudo-topics, and that lack boundary control and result traceability. This would improve the accuracy, reliability of evidence, controllability of boundaries and interpretability of results in topic discovery. Summary of the Invention

[0005] The purpose of this application is to provide a multi-dimensional topic discovery method, system, device, medium, and product based on prior information and boundary awareness, which can perform evidence calibration, boundary-aware retrieval, and content discovery on the true expression of user-given prior topic information in text corpora, thereby improving the accuracy, reliability of evidence, controllability of boundaries, and interpretability of results in topic discovery.

[0006] To achieve the above objectives, this application provides the following solution: Firstly, this application provides a multi-dimensional topic discovery method based on prior information and boundary awareness, including: Obtain the target text corpus set and the corresponding prior topic information set; each prior topic information set includes a core word set, an extended word set, and a negative boundary word set; The target text corpus is segmented to obtain a set of text units; Construct a positive topic prototype and a negative boundary prototype for each prior topic in the prior topic information set; Based on the set of text units, the topic support level of each prior topic information in the prior topic information set is determined; the topic support level includes sufficient evidence level, moderate evidence level, weak evidence level, and insufficient evidence level. Prior topic information with a topic support level of insufficient evidence is marked as insufficient evidence prior topic information, and prior topic information with a topic support level of sufficient evidence, moderate evidence, and weak evidence is marked as valid prior topic information. For each valid prior topic information in the prior topic information set, based on the corresponding positive topic prototype and negative boundary prototype, soft evidence text paragraph retrieval is performed on the text unit set to obtain a candidate text evidence paragraph set corresponding to each valid prior topic information in the prior topic information set; the soft evidence text paragraph retrieval includes word literal matching retrieval, semantic similarity retrieval, and negative boundary constraint retrieval; Each candidate text evidence paragraph in the candidate text evidence paragraph set corresponding to each valid prior topic information in the prior topic information set is filtered to obtain the topic discovery result for each valid prior topic information; the topic discovery result includes topic support level, topic word set, topic text evidence paragraph, and boundary risk warning.

[0007] Secondly, this application provides a multi-dimensional topic discovery system based on prior information and boundary awareness, including: The data acquisition module is used to acquire the target text corpus set and the corresponding prior topic information set; each prior topic information set includes a core word set, an extended word set, and a negative boundary word set. The text segmentation module is used to segment the target text corpus to obtain a set of text units. The prototype building module is used to build the positive topic prototype and negative boundary prototype for each prior topic in the prior topic information set. The topic support level determination module is used to determine the topic support level of each prior topic information in the prior topic information set based on the set of text units; the topic support level includes sufficient evidence level, moderate evidence level, weak evidence level, and insufficient evidence level; The topic labeling module is used to label prior topic information with a topic support level of insufficient evidence as insufficient evidence prior topic information, and to label prior topic information with a topic support level of sufficient evidence, moderate evidence, and weak evidence as valid prior topic information. The soft evidence retrieval module is used to perform soft evidence text paragraph retrieval on the text unit set for each valid prior topic information in the prior topic information set, based on the corresponding positive topic prototype and negative boundary prototype, to obtain a candidate text evidence paragraph set corresponding to each valid prior topic information in the prior topic information set; the soft evidence text paragraph retrieval includes word literal matching retrieval, semantic similarity retrieval, and negative boundary constraint retrieval. The topic discovery result generation module is used to filter all candidate text evidence paragraphs in the candidate text evidence paragraph set corresponding to each valid prior topic information in the prior topic information set, and obtain the topic discovery result for each valid prior topic information; the topic discovery result includes topic support level, topic word set, topic text evidence paragraph, and boundary risk warning.

[0008] Thirdly, this application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the multi-dimensional topic discovery method based on prior information and boundary awareness described above.

[0009] Fourthly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the multi-dimensional topic discovery method based on prior information and boundary awareness described above.

[0010] Fifthly, this application provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the multi-dimensional topic discovery method based on prior information and boundary awareness described above.

[0011] According to the specific embodiments provided in this application, this application has the following technical effects: This application provides a multi-dimensional topic discovery method, system, device, medium, and product based on prior information and boundary awareness. By acquiring a target text corpus and a corresponding set of prior topic information, and segmenting the target text corpus to obtain a set of text units, it solves the problem that traditional topic modeling cannot handle the conceptual dimensions that researchers focus on beforehand, achieving structured input of prior topic information and preprocessing of the corpus. By constructing positive topic prototypes and negative boundary prototypes for each prior topic information in the prior topic information set, it addresses the problem that seed-word-guided methods rely solely on positive seed words for semantic expansion and lack effective control over topic boundaries, achieving bidirectional constraints on topic boundaries and effective differentiation of semantic directions. Furthermore, by determining the topic support level of each prior topic information in the prior topic information set based on the set of text units, prior topic information with a support level of insufficient evidence is marked as insufficient evidence prior topic information, while prior topic information with support levels of sufficient evidence, moderate evidence, and weak evidence is marked as valid prior topic information, solving the problem of... The seed-word-guided method lacks a mechanism to determine whether a topic truly exists in the corpus and is prone to generating pseudo-topics. This method effectively verifies topic discoverability and filters out pseudo-topics. For each valid prior topic in the prior topic information set, based on the corresponding positive topic prototype and negative boundary prototype, soft evidence text paragraph retrieval is performed on the text unit set, including word-based matching retrieval, semantic similarity retrieval, and negative boundary constraint retrieval. This yields a set of candidate text evidence paragraphs corresponding to each valid prior topic in the prior topic information set, solving the problem of traditional hard rule retrieval easily deleting cross-topic expressions and achieving flexible identification and retention of topic-related paragraphs in complex texts. By filtering all candidate text evidence paragraphs in the candidate text evidence paragraph set corresponding to each valid prior topic in the prior topic information set, topic discovery results including topic support level, topic word set, topic text evidence paragraphs, and boundary risk warnings are obtained. This solves the problem of traditional methods having untraceable results and being unable to return to specific text paragraphs for verification, achieving interpretability and reliability of topic discovery results. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1A flowchart illustrating a multi-dimensional topic discovery method based on prior information and boundary awareness, provided in an embodiment of this application; Figure 2 A schematic diagram of the functional modules of a topic discovery system for prior topic information provided in an embodiment of this application; Figure 3 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0014] First, some technical terms involved in the embodiments of this application will be introduced.

[0015] Given prior topic information or seed words (usually in the form of words or phrases) from the user, this mechanism identifies the content most relevant to these prior topics from a large volume of text and determines whether there is sufficient evidence to support these topics in the corpus. It doesn't simply perform unsupervised topic discovery or keyword matching; rather, it aims to establish a computationally achievable, traceable, and controllable boundary-based topic identification mechanism between "prior guidance" and "textual evidence."

[0016] The basic principles in related technologies can be summarized into four levels.

[0017] First, the method uses multiple layers of seed words to define each prior topic. A topic is no longer represented by a single keyword, but rather by core words, extended words, and negative boundary words. Core words define the main idea of ​​the topic, extended words broaden the recall scope, and negative boundary words exclude adjacent topics and generalize expressions. The significance of this is that topics are not defined solely by "similarity," but by both "what should be approached" and "what should be avoided." For example, "supply chain management" should not only be close to "procurement, inventory, logistics, and suppliers," but also differentiated from adjacent or unrelated directions such as "R&D design, production line, equipment maintenance, and financial management."

[0018] Second, the method first determines whether a topic is discoverable in the current corpus. That is, it doesn't simply force a topic output based on a given seed word. The system checks whether the core word appears, whether the extended words cover the topic, and whether there are enough semantically similar text units. If a topic is almost entirely absent from the text, it's marked as insufficient evidence, rather than forcibly generating a seemingly related set of topic words. This step avoids generating "pseudo-topics," enabling the method to distinguish between "topics that truly exist in the corpus" and "topics that theoretically exist but lack evidence in the text."

[0019] Third, the method uses soft evidence retrieval to find sentences or paragraphs that support a specific topic. It doesn't simply require paragraphs to contain only the current topic's keywords and exclude other topics; instead, it comprehensively assesses the relevance and purity of a text unit to various topics. This allows for better handling of complex texts, especially cross-topic expressions common in annual reports, policy documents, and academic abstracts. For example, a paragraph might simultaneously mention "supply chain platform, procurement, production, inventory, and logistics," encompassing both supply chain management information and information on manufacturing and cross-link collaboration. Traditional hard rules might exclude it, but soft evidence retrieval will retain it and mark its topic affiliation strength or cross-topic attributes.

[0020] Fourth, the method filters keywords from high-quality evidence paragraphs and performs boundary-aware ranking. Whether a word can become a keyword for a particular topic depends not only on its frequency of occurrence but also on whether it primarily appears in evidence paragraphs related to that topic, whether it is semantically close to the current topic, whether it is closer to the current topic than other topics, and whether it carries a high boundary risk. Words like "system," "platform," "data," and "management," while potentially frequent, are given lower weight due to their broad cross-topic distribution and weak distinguishability. Conversely, more focused and scenario-specific words like "inventory optimization," "supplier collaboration," "order management," and "logistics delivery" are more likely to be retained as keywords.

[0021] Therefore, the overall logic of this method is as follows: first, define the topic boundaries using structured seed words; then, determine whether there is sufficient textual evidence for the topic; then, retrieve evidence paragraphs from the corpus that can support the topic; and finally, select the main content with relevance, distinctiveness, and boundary stability from these evidence paragraphs.

[0022] In summary, the method proposed in this application aims to improve three capabilities of prior-based text topic discovery: first, evidence calibration capability, i.e., determining whether a topic is truly supported by the corpus; second, boundary control capability, i.e. reducing confusion caused by adjacent topics and generalized words; and third, result traceability capability, i.e., each discovered topic can be traced back to specific text paragraphs to find evidence.

[0023] From an application perspective, this method is particularly suitable for tasks where researchers already have a theoretical framework or classification dimensions, but need to identify how these dimensions are expressed from a large amount of text. Examples include identifying digital application scenarios in corporate annual reports, analyzing governance themes in policy texts, classifying research themes in literature reviews, and identifying technological directions in patent texts. It can help researchers transform abstract prior concepts into computable textual evidence and keyword sets, thereby serving subsequent theme analysis, text classification, evidence tracking, and quantitative evaluation.

[0024] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0025] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0026] In one exemplary embodiment, such as Figure 1 As shown, a multi-dimensional topic discovery method based on prior information and boundary awareness is provided. This method is executed by a computer device, specifically by a terminal or server alone, or by both a terminal and a server, and includes the following steps 101 to 107. Wherein: Step 101: Obtain the target text corpus set and the corresponding prior topic information set; each prior topic information set includes a core word set, an extended word set, and a negative boundary word set.

[0027] Step 102: Perform text segmentation on the target text corpus to obtain a set of text units.

[0028] Step 103: Construct the positive topic prototype and negative boundary prototype for each prior topic information in the prior topic information set.

[0029] Step 104: Based on the set of text units, determine the topic support level of each prior topic information in the prior topic information set; the topic support level includes sufficient evidence level, moderate evidence level, weak evidence level, and insufficient evidence level.

[0030] Step 105: Mark the prior topic information in the prior topic information set that has a topic support level of insufficient evidence as insufficient evidence prior topic information, and mark the prior topic information that has a topic support level of sufficient evidence, average evidence, and weak evidence as valid prior topic information.

[0031] Step 106: For each valid prior topic information in the prior topic information set, based on the corresponding positive topic prototype and negative boundary prototype, perform soft evidence text paragraph retrieval on the text unit set to obtain a candidate text evidence paragraph set corresponding to each valid prior topic information in the prior topic information set; the soft evidence text paragraph retrieval includes word literal matching retrieval, semantic similarity retrieval, and negative boundary constraint retrieval.

[0032] Step 107: Filter all candidate text evidence paragraphs in the candidate text evidence paragraph set corresponding to each valid prior topic information in the prior topic information set to obtain the topic discovery result for each valid prior topic information; the topic discovery result includes topic support level, topic word set, topic text evidence paragraph, and boundary risk warning.

[0033] By implementing steps 101 to 107 above, this application can, under the condition of given prior topic information, define topic boundaries through structured seeds, verify the actual existence of topics in the corpus, flexibly retrieve textual evidence supporting the topics, and extract representative topic words in a boundary-aware manner. This enables evidence calibration, boundary control, and result traceability in the topic discovery process. It effectively solves the technical problems that traditional unsupervised topic modeling cannot correspond to pre-focused dimensions and that seed-guided methods are prone to generating pseudo-topics due to a lack of evidence judgment and boundary control. This significantly improves the accuracy, reliability, and interpretability of topic discovery.

[0034] In another exemplary embodiment of this application, a positive topic prototype and a negative boundary prototype are constructed for each prior topic information in the prior topic information set, specifically including: For each prior topic information in the prior topic information set, each word in the corresponding core word set, extended word set, and negative boundary word set is input into the pre-trained text embedding model to obtain the core word vector representation set, extended word vector representation set, and negative boundary word vector representation set corresponding to each prior topic information. The positive topic prototype and negative boundary prototype of each prior topic in the prior topic information set are constructed using the following formulas: ; ; in, Indicates the first A positive topic prototype based on prior topic information; Represents the normalization function; Indicates the first A set of core words for a priori topic information; Indicates the weight coefficient of the core words; Indicates the first The core word vector representation corresponding to each core word; Indicates the first An expanded set of terms for prior topic information; Indicates the weight coefficient of the expanded words; Indicates the first The extended word vector representation corresponding to each extended word; Indicates the first A negative boundary prototype of prior subject information; Indicates the first A set of negative boundary words for prior topic information; The weight coefficient representing negative boundary words; Indicates the first The negative boundary word vector representation corresponding to each negative boundary word; Indicates the first A set of core words for a priori topic information; This indicates that, except for the first element, the set of prior subject information contains... The weight coefficients of the core words of all prior topic information other than the first prior topic information.

[0035] In another exemplary embodiment of this application, based on the set of text units, the topic support level of each prior topic information in the prior topic information set is determined, specifically including: The core word coverage of each prior topic information is calculated using the following formulas: ; The expanded word coverage of each prior topic information is calculated using the following formulas: ; The semantic evidence coverage of each prior topic information is calculated using the following formulas: ; The topic support for each prior topic information is calculated using the following formulas: ; in, Indicates the first The coverage of core keywords in prior topic information; Indicates the first One core keyword; Indicates the first A set of core words for a priori topic information; Indicates the first The core words in the target text corpus Word frequency in; Indicates the first Expansion word coverage of prior topic information; Indicates the first One extended term; Indicates the first A set of negative boundary words for prior topic information; Indicates the first The extended words in the target text corpus Word frequency in; Indicates the first The semantic evidence coverage of prior topic information; Indicates the first Each text unit; Represents a collection of text units; Indicates an indicator function; Indicates the first The text unit vector corresponding to each text unit With the A positive topic prototype of prior topic information Cosine similarity; This indicates a preset semantic similarity threshold; Indicates the first The topic support of each prior topic information; , and These represent the weighting coefficients for core word coverage, extended word coverage, and semantic evidence coverage, respectively. If the first If the topic support of the first prior topic information is less than the first preset support threshold, then the... The prior topic information has a topic support level of insufficient evidence. If the first If the topic support of the first prior topic information is greater than or equal to the first preset support threshold and less than the second preset support threshold, then the first... The prior topic information has a weak level of support for the evidence. If the first If the topic support of the first prior topic information is greater than or equal to the second preset support threshold and less than the third preset support threshold, then the first... The thematic support level of the prior thematic information is the general level of evidence. If the first If the topic support of the first prior topic information is greater than or equal to the preset threshold of the third support, then the first... The thematic support level of each prior topic information is the level of sufficient evidence.

[0036] In another exemplary embodiment of this application, for each valid prior topic information in the prior topic information set, based on the corresponding positive topic prototype and negative boundary prototype, soft evidence text paragraph retrieval is performed on the text unit set to obtain a candidate text evidence paragraph set corresponding to each valid prior topic information in the prior topic information set, specifically including: The word-literal matching score between any text unit in the text unit set and each valid prior topic information is calculated using the following formulas: ; The semantic similarity score and negative boundary score between any text unit in the text unit set and each valid prior topic information are calculated using the following formulas: ; ; The comprehensive evidentiary relevance between any text unit in the text unit set and each valid prior topic information is calculated using the following formulas: ; The soft purity score of any text unit in the text unit set relative to each valid prior topic information is calculated using the following formulas: ; in, Indicates the first The text unit and the first A word literal matching score for each valid prior topic information; Indicates the first One core keyword; Indicates the first A set of core words for a valid prior topic information; Indicates the weight coefficient of the core words; Indicates the first The first core keyword and the first BM25 word literal matching strength of each text unit; Indicates the first One extended term; Indicates the first An expanded set of words containing valid prior topic information; Indicates the weight coefficient of the expanded words; Indicates the first The first extended word and the first BM25 word literal matching strength of each text unit; Indicates the first One negative boundary word; Indicates the first A set of negative boundary words containing valid prior topic information; The weight coefficient representing negative boundary words; Indicates the first The negative boundary word and the first BM25 word literal matching strength of each text unit; Indicates the first The text unit and the first The semantic similarity score of each valid prior topic information; Indicates the first The text unit vector corresponding to each text unit With the A positive topic prototype of valid prior topic information Cosine similarity; Indicates the first The text unit and the first Negative boundary scores for each valid prior topic information; Indicates the first The text unit vector corresponding to each text unit With the Negative boundary prototype of valid prior topic information Cosine similarity; Indicates the first The text unit and the first The overall evidentiary relevance of each valid prior topic information; , and These represent the weight coefficients of the word literal matching score, the semantic similarity score, and the negative boundary score, respectively, all of which are greater than zero; Indicates the first The text unit and the first The soft purity score of each valid prior topic information; Indicates the first The text unit and the first The overall evidentiary relevance of each valid prior topic information; Indicates the first The background topic score for each text unit is used to represent generalized or irrelevant text. It can be set as a constant or as a function of the proportion of generalized words. Text units whose overall evidence relevance exceeds a preset relevance threshold are selected as the first... Add the candidate text evidence paragraphs containing the first valid prior topic information to the first... A set of candidate textual evidence paragraphs containing valid prior topic information.

[0037] In another exemplary embodiment of this application, all candidate text evidence paragraphs in the candidate text evidence paragraph set corresponding to each valid prior topic information in the prior topic information set are filtered to obtain the topic discovery result for each valid prior topic information, specifically including: For each valid prior topic information in the prior topic information set, all candidate text evidence paragraphs in the candidate text evidence paragraph set are added to the topic text evidence paragraph set corresponding to each valid prior topic information if the overall evidence relevance exceeds a preset relevance threshold and the soft purity score exceeds a preset soft purity score threshold. Candidate text evidence paragraphs with an overall evidence relevance exceeding a preset relevance threshold and a soft purity score not exceeding a preset soft purity score threshold are added to the cross-topic text evidence paragraph set corresponding to each valid prior topic information. Candidate words are extracted from the set of topic text evidence paragraphs corresponding to each valid prior topic information to obtain the candidate word set corresponding to each valid prior topic information. Each candidate word set corresponding to a valid prior topic is sorted and filtered using boundary awareness to obtain the topic discovery results corresponding to each valid prior topic.

[0038] In another exemplary embodiment of this application, boundary-aware sorting and filtering are performed on the candidate word set corresponding to each valid prior topic information to obtain the topic discovery result corresponding to each valid prior topic information, specifically including: The weighted occurrence intensity of each candidate word in the candidate word set corresponding to each valid prior topic information in the corresponding set of topic text evidence paragraphs of the valid prior topic information is calculated using the following formulas: ; The topic discriminative scores of each candidate word in the candidate word set corresponding to each valid prior topic information are calculated using the following formulas: ; The evidence score of each candidate word in the candidate word set corresponding to each valid prior topic information is calculated using the following formulas: ; The semantic relevance, cross-topic marginal value, and boundary risk of each candidate word in the candidate word set corresponding to each valid prior topic information are calculated using the following formulas: ; ; ; The final ranking score of each candidate word in the candidate word set corresponding to each valid prior topic information is calculated using the following formulas: ; in, Indicates the first The candidate word set corresponding to the _th valid prior topic information is the _th The candidate word in the first The weighted occurrence intensity of the thematic text evidence paragraphs in a set of valid prior thematic information; Indicates the first A set of thematic textual evidence paragraphs containing valid prior thematic information; Indicates the first The text unit and the first The soft purity score of each valid prior topic information; Indicates the first The candidate word set corresponding to the _th valid prior topic information is the _th The candidate word in the first Word frequency in each text unit; Indicates the first The candidate word set corresponding to the _th valid prior topic information is the _th The candidate word is relative to the first The topic discrimination score of each valid prior topic information; Indicates the first The candidate word set corresponding to the _th valid prior topic information is the _th The candidate word and the first A collection of thematic text evidence paragraphs containing valid prior thematic information. BM25 literal matching strength; Indicates the smoothing coefficient; Indicates the first The candidate word set corresponding to the _th valid prior topic information is the _th The candidate word and the first A collection of thematic text evidence paragraphs containing valid prior thematic information. BM25 literal matching strength; Indicates the first The candidate word set corresponding to the _th valid prior topic information is the _th The candidate word is relative to the first Evidence score for each valid prior topic information; This represents a balance coefficient between the weighted intensity and thematic distinctiveness; Indicates the first The candidate word set corresponding to the _th valid prior topic information is the _th The candidate word and the first The semantic relevance of a valid prior topic information to a positive topic prototype; Indicates the first The candidate word set corresponding to the _th valid prior topic information is the _th Candidate word vector representation of candidate words With the A positive topic prototype of valid prior topic information Cosine similarity; Indicates the first The candidate word set corresponding to the _th valid prior topic information is the _th The candidate word and the first The cross-topic marginal value of a valid prior topic information positive topic prototype; Indicates the first The candidate word set corresponding to the _th valid prior topic information is the _th Candidate word vector representation of candidate words With the A positive topic prototype of valid prior topic information Cosine similarity; Indicates the first The candidate word set corresponding to the _th valid prior topic information is the _th The candidate word and the first Boundary risk of a valid prior topic information positive topic prototype; Indicates the first The topic distribution entropy of candidate words; Indicates the first The candidate word set corresponding to the _th valid prior topic information is the _th The candidate word is relative to the first The final ranking score of each valid prior topic information; , and These represent the weighting coefficients for semantic relevance, cross-topic marginal value, and boundary risk, respectively. Based on the final ranking score, topic words are selected from the candidate word set corresponding to each valid prior topic information using a preset selection method and added to the topic word set corresponding to each valid prior topic information, while simultaneously generating boundary risk warnings; the preset selection method includes sorting each candidate word in the candidate word set corresponding to each valid prior topic information from high to low according to the final ranking score, and selecting the first preset number of candidate words as topic words; or, selecting candidate words in the candidate word set corresponding to each valid prior topic information whose final ranking score exceeds a preset threshold as topic words; The topic discovery results for each valid prior topic information are the topic term set, boundary risk warning, topic support level, and topic text evidence paragraph corresponding to each valid prior topic information.

[0039] The following example illustrates this application using a specific multi-dimensional topic discovery process based on prior information and boundary awareness.

[0040] Given a target text corpus D, divide it into a set of text units U. Text units can be sentences, paragraphs, or clauses / sub-clauses. Clauses / sub-clauses are grammatical units between words / phrases and complete sentences. For complex texts such as corporate annual reports, policy texts, industry reports, and paper abstracts, it is recommended to use paragraphs or paragraphs enhanced with clauses.

[0041] D = {d1, d2, ..., d} N}; U = {u1, u2, ..., u} M}; Among them, d1, d2 and d N Let u1, u2, and u3 represent the first, second, and Nth subsets of the target text corpus, respectively; N represents the total number of subsets of the target text corpus in the target text corpus set; u1, u2, and u3 represent the subsets of the target text corpus. M These represent the first, second, and Mth text units, respectively; M represents the total number of text units in the text unit set; where the target text corpus subset can be a batch of text corpora of the same category, such as dividing the annual reports of a certain region into batches of 10-year units, and then performing text unit segmentation and theme extraction on each batch of annual reports.

[0042] Given a prior set of topic information C, for each topic C i It consists of three types of seed words: core seed words (i.e., core words, used to define the main theme, emphasizing high precision and strong representativeness), extended seed words (i.e. extended words, used to expand the recall scope, covering synonyms and domain expressions), and negative boundary seed words (i.e. negative boundary words, used to control the topic boundary, penalizing adjacent topic words and generalized words).

[0043] C = {C1, C2, ..., C} K}; ={ , , }; For example, for the topic of supply chain management, the following three types of seed keywords can be set: ={Supply Chain, Procurement, Inventory, Logistics, Suppliers}; ={Supply chain platform, order collaboration, inventory optimization, warehousing and distribution, supplier collaboration}; ={Research and development design, production line, equipment operation and maintenance, financial management, customer service}.

[0044] For each topic As shown in Table 1 below, the topic discovery results output by the method include topic support level, topic words, evidence paragraphs, and boundary risk warnings. This method does not force each seed word to find a topic in the text corpus; instead, it first determines whether there is sufficient evidence for the topic in the current corpus.

[0045] Table 1. Topic Discovery Results

[0046] In one exemplary embodiment, a multi-dimensional topic discovery method based on prior information and boundary awareness is provided, comprising the following steps: Step 1: Constructing multi-level seed term priors.

[0047] Seed-word-guided topic discovery typically assigns only one or a few positive seed words to each topic. This can easily lead to insufficient recall and a lack of boundary control over adjacent topics and generalization words. Therefore, the first core component defines each prior topic information as a three-layer structure.

[0048] ={ , , }; (1) Keyword vector representation.

[0049] For each seed word, a vector representation is obtained using a pre-trained text embedding model (e.g., BGE, text2vec, SimCSE, Sentence-BERT, Chinese-RoBERTa sentence embedding). If the seed word is a phrase, such as a supply chain collaboration platform, it is recommended to encode the complete phrase directly, rather than simply averaging the split words.

[0050] ; ; You can also use definition enhancement forms to make the vectors of seed words more closely reflect the current topic context.

[0051] ; For example, “inventory optimization” is coded as “supply chain management related concept: inventory optimization”, rather than being coded as “inventory optimization” in isolation.

[0052] in, ; This indicates the definition of an enhanced template function. It involves embedding the isolated seed word k into the topic. In the context of the topic, this makes the generated vectors more semantically relevant to the current topic.

[0053] (2) Construct a positive theme prototype.

[0054] Prior subject information Build positive topic prototypes using core seed words and extended seed words. The weight of core seed keywords should be higher than that of expanded seed keywords.

[0055] ; in, > >0, for example, can be set to =1.0, =0.6 indicates that the core seed words mainly define the theme, while the extended seed words mainly expand the semantic coverage.

[0056] (3) Construct a negative boundary prototype.

[0057] Construct a negative boundary prototype using negative seed words and core seed words from other themes. This indicates the semantic direction that the current topic should stay away from.

[0058] ; For supply chain management, negative boundaries may include areas such as R&D design, manufacturing, operation and maintenance services, and financial management.

[0059] Step 2: Topic Discoverability Assessment. This step determines whether the current corpus truly contains topic information that matches prior knowledge. Relevant textual evidence. If there is insufficient evidence for a particular topic, the topic keywords should not be forced to appear; instead, they should be marked as insufficient evidence.

[0060] (1) Coverage of core words.

[0061] Core word coverage measures the direct literal evidence of core seed words in the entire corpus.

[0062] ; like The low score indicates a lack of direct linguistic evidence for the topic.

[0063] (2) Coverage of extended words.

[0064] Expanded word coverage measures the extent to which synonyms, alternative expressions, or domain expressions are covered in the corpus.

[0065] ; For example, if "supply chain collaboration" appears infrequently, but "inventory optimization," "logistics and distribution," and "order management" appear frequently, then this topic may still have high support.

[0066] (3) Semantic evidence coverage.

[0067] For each text unit u, first calculate its vector representation, then calculate its relationship with the positive topic prototype. The cosine similarity.

[0068] ; ; Semantic evidence coverage indicates how many text units are semantically close to prior topic information. .

[0069] ; To facilitate comparisons between corpora of different sizes, normalization can also be used.

[0070] .

[0071] in, It is a semantic similarity threshold, which controls the strictness of semantic evidence screening. A higher value retains less text, but usually results in higher precision; A lower precision / recall ratio recalls more relevant expressions, but may include weakly relevant or generalized content. Essentially, it controls the precision-recall balance in semantic evidence recognition. A more accurate definition is the minimum semantic similarity threshold required for a text unit to be recognized as potential semantic evidence for a given prior topic.

[0072] (4) Overall score of topic support.

[0073] The core word coverage, extended word coverage, and normalized semantic evidence coverage are combined to form topic support.

[0074] ; ω1+ω2+ω3=1; Weights can be set according to the task, for example, ω1=0.4, ω2=0.3, ω3=0.3. Then, based on... The topics are categorized into four levels: sufficient evidence, moderate evidence, weak evidence, and insufficient evidence, as shown in Table 2 below.

[0075] Table 2 Topic Support Levels

[0076] Step 3: Soft evidence paragraph retrieval. Traditional methods often use hard rules to find topic indicator sentences, such as requiring sentences to contain the current topic term and not other topic terms. This can easily lead to the accidental deletion of valuable cross-topic paragraphs in complex texts. Therefore, this step employs soft evidence retrieval, making soft judgments on the relevance and purity of each topic based on text units.

[0077] (1) Literal matching score.

[0078] For text unit u, calculate its relationship with prior topic information. The word-to-word matching score. A higher score is awarded for a core word match, a lower score for an extended word match, and a lower score for a negative word match.

[0079] ; Among them, BM25 is used to measure the strength of literal evidence between a term and a text unit.

[0080] ; Here, |u| is suggested to be defined as the number of valid tokens after Chinese word segmentation, domain phrase merging, punctuation filtering, and stop word processing; avglen is the average length of all text units. Seed word The less common the inverse document in the entire corpus, the higher the frequency. This indicates the degree of word frequency saturation, typically set to 1.2-2.0; This indicates the control of the normalization intensity, which is generally taken as 0.75. The above is the general range of parameter values, which can be adjusted according to the results of specific data experiments.

[0081] (2) Semantic similarity and negative boundary score.

[0082] Semantic similarity is used to measure whether the overall semantics of a text unit are close to the current topic; negative boundary score is used to measure whether a text unit is close to the direction that should be excluded from the current topic.

[0083] ; ; (3) Overall relevance of evidence By combining word-meaning matching, semantic similarity, and negative boundary scores, the evidentiary relevance of text unit u to valid prior topic information is obtained.

[0084] ; Where a, b, and c are all greater than 0. For example, we can set a = 0.4, b = 0.4, and c = 0.2.

[0085] (4) Calculation of subject purity (soft purity fraction).

[0086] To determine whether a text unit primarily supports the current topic, soft topic purity is calculated.

[0087] ; in, This is the background topic score, used to represent generalized or irrelevant text. It can be set to a constant or a function of the proportion of generalized words.

[0088] (5) Selection of evidence paragraphs.

[0089] according to and The text units are divided into strong-topic text evidence paragraphs, cross-topic text evidence paragraphs, and weak-topic text evidence paragraphs, as shown in Table 3 below.

[0090] Table 3. Textual Evidence Paragraph Types

[0091] .

[0092] The value is not a fixed, universal value because the cosine similarity distribution differs across different embedding models, corpora, and topics. It needs to be determined based on the embedding model, corpus, and topic used. , This patent can adopt a scheme of "adaptive topic relevance and relatively uniform soft topic purity".

[0093] express The set of background relevance scores, not a single score, is defined as: ; in, Indicating targeting A collection of background text constructed. For example, It is "supply chain management". This can include: text on other topics, such as R&D design, manufacturing, and operation and maintenance services; text containing negative boundary words; non-target topic paragraphs randomly selected from the corpus; and generalized text that is adjacent to but does not belong to supply chain management.

[0094] Step 4: Boundary-aware keyword sorting.

[0095] This step is used to extract... The process involves filtering multiple keywords. The goal is not simply to find high-frequency words, but rather to identify words or phrases that frequently appear in the current topic's evidence, are semantically close to the topic, are distinguishable from other topics, and have low boundary risks.

[0096] (1) A set of candidate words or phrases.

[0097] First, candidate words and phrases are extracted from the corpus to form a set of candidate words or phrases, V.

[0098] V={w1,w2,…,w L}; Candidate words or phrases can come from Chinese word segmentation results, noun phrases, domain dictionaries, and high-frequency phrase extraction results.

[0099] (2) Evidence score.

[0100] The weighted intensity of a candidate word or phrase in a paragraph containing high-purity evidence on the current topic is defined as: ; The higher the score, the more frequently the candidate words appear in the evidence paragraphs that represent the current topic.

[0101] (3) Thematic differentiation score.

[0102] The discriminative score is used to determine whether candidate words appear primarily in evidence paragraphs on the current topic, rather than appearing widely in paragraphs on other topics.

[0103] ; ; (4) Evidence-based keyword scores.

[0104] The evidence score is synthesized by combining the universality within the subject and the distinctiveness of the subject.

[0105] ; Where 0 < α < 1. If more emphasis is placed on the frequency in high-purity paragraphs, α can be set to 0.6; if more emphasis is placed on thematic differentiation, α can be set to 0.4.

[0106] (5) Semantic relevance, cross-topic margins and boundary risks.

[0107] Semantic relevance measures the similarity between candidate words and current topic prototypes.

[0108] ; Cross-topic marginal measures whether candidate words are closer to the current topic than other topics.

[0109] ; Boundary risk is determined by the similarity and generalization between candidate words and negative boundary prototypes.

[0110] ; in, It can be measured using topic distribution entropy.

[0111] ; ; If a word is similar across multiple topics, such as "platform," "system," "data," and "management," its topic distribution will be more even, its entropy will be higher, and its generalization will be greater.

[0112] (6) Final keyword score.

[0113] The final ranking score of candidate words is obtained by combining evidence score, semantic relevance, cross-topic margin, and boundary risk.

[0114] ; β1+β2+β3+β4=1; β1 can be set to 0.4, β2 to 0.25, β3 to 0.2, and β4 to 0.15. The top-K candidate words or those exceeding the thresholds are ultimately selected as the topic words.

[0115] If there is insufficient topic support, no formal topic terms will be output.

[0116] Based on the same inventive concept, this application also provides a multi-dimensional topic discovery system based on prior information and boundary awareness for implementing the aforementioned multi-dimensional topic discovery method based on prior information and boundary awareness. The solution provided by this system is similar to the implementation scheme described in the above method. Therefore, the specific limitations of one or more embodiments of the multi-dimensional topic discovery system based on prior information and boundary awareness provided below can be found in the limitations of the multi-dimensional topic discovery method based on prior information and boundary awareness described above, and will not be repeated here.

[0117] In one exemplary embodiment, such as Figure 2 As shown, a multi-dimensional topic discovery system based on prior information and boundary awareness is provided, including: The data acquisition module 201 is used to acquire the target text corpus set and the corresponding prior topic information set; each prior topic information set includes a core word set, an extended word set, and a negative boundary word set.

[0118] The text segmentation module 202 is used to segment the target text corpus to obtain a set of text units.

[0119] The prototype building module 203 is used to build the positive topic prototype and the negative boundary prototype of each prior topic information in the prior topic information set.

[0120] The topic support level determination module 204 is used to determine the topic support level of each prior topic information in the prior topic information set based on the set of text units; the topic support level includes sufficient evidence level, moderate evidence level, weak evidence level and insufficient evidence level.

[0121] The topic labeling module 205 is used to label prior topic information in the prior topic information set that has a topic support level of insufficient evidence as insufficient evidence prior topic information, and to label prior topic information that has a topic support level of sufficient evidence, moderate evidence, and weak evidence as valid prior topic information.

[0122] The soft evidence retrieval module 206 is used to perform soft evidence text paragraph retrieval on the text unit set for each valid prior topic information in the prior topic information set, based on the corresponding positive topic prototype and negative boundary prototype, to obtain a candidate text evidence paragraph set corresponding to each valid prior topic information in the prior topic information set; the soft evidence text paragraph retrieval includes word literal matching retrieval, semantic similarity retrieval, and negative boundary constraint retrieval.

[0123] The topic discovery result generation module 207 is used to filter all candidate text evidence paragraphs in the candidate text evidence paragraph set corresponding to each valid prior topic information in the prior topic information set, and obtain the topic discovery result for each valid prior topic information; the topic discovery result includes topic support level, topic word set, topic text evidence paragraph and boundary risk warning.

[0124] In one exemplary embodiment, a computer device is provided, which may be a server or a terminal, and its internal structure diagram may be as follows. Figure 3 As shown, this computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operating system and computer programs in the non-volatile storage media to run. The database stores multi-dimensional topic discovery processing data based on prior information and boundary awareness. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network connection. When the computer program is executed by the processor, it implements a multi-dimensional topic discovery method based on prior information and boundary awareness.

[0125] Those skilled in the art will understand that Figure 3 The structures shown are merely block diagrams of some structures related to the present application and do not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than shown in the figures, or combine certain components, or have different component arrangements. In an exemplary embodiment, a computer device is provided, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0126] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0127] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0128] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0129] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0130] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0131] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0132] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A multi-dimensional topic discovery method based on prior information and boundary awareness, characterized in that, include: Obtain the target text corpus set and the corresponding prior topic information set; each prior topic information set includes a core word set, an extended word set, and a negative boundary word set; The target text corpus is segmented to obtain a set of text units; Construct a positive topic prototype and a negative boundary prototype for each prior topic in the prior topic information set; Based on the set of text units, determine the topic support level of each prior topic information in the prior topic information set; The subject matter support levels include the level of sufficient evidence, the level of moderate evidence, the level of weak evidence, and the level of insufficient evidence. Prior topic information with a topic support level of insufficient evidence is marked as insufficient evidence prior topic information, and prior topic information with a topic support level of sufficient evidence, moderate evidence, and weak evidence is marked as valid prior topic information. For each valid prior topic information in the prior topic information set, based on the corresponding positive topic prototype and negative boundary prototype, soft evidence text paragraph retrieval is performed on the text unit set to obtain a candidate text evidence paragraph set corresponding to each valid prior topic information in the prior topic information set; the soft evidence text paragraph retrieval includes word literal matching retrieval, semantic similarity retrieval, and negative boundary constraint retrieval; Each candidate text evidence paragraph in the candidate text evidence paragraph set corresponding to each valid prior topic information in the prior topic information set is filtered to obtain the topic discovery result for each valid prior topic information; the topic discovery result includes topic support level, topic word set, topic text evidence paragraph, and boundary risk warning.

2. The multi-dimensional topic discovery method based on prior information and boundary awareness according to claim 1, characterized in that, Construct a positive topic prototype and a negative boundary prototype for each prior topic in the prior topic information set, specifically including: For each prior topic information in the prior topic information set, each word in the corresponding core word set, extended word set, and negative boundary word set is input into the pre-trained text embedding model to obtain the core word vector representation set, extended word vector representation set, and negative boundary word vector representation set corresponding to each prior topic information. The positive topic prototype and negative boundary prototype of each prior topic in the prior topic information set are constructed using the following formulas: ； ； in, Indicates the first A positive topic prototype based on prior topic information; Represents the normalization function; Indicates the first A set of core words for a priori topic information; Indicates the weight coefficient of the core words; Indicates the first The core word vector representation corresponding to each core word; Indicates the first An expanded set of terms for prior topic information; Indicates the weight coefficient of the expanded words; Indicates the first The extended word vector representation corresponding to each extended word; Indicates the first A negative boundary prototype of prior subject information; Indicates the first A set of negative boundary words for prior topic information; The weight coefficient representing negative boundary words; Indicates the first The negative boundary word vector representation corresponding to each negative boundary word; Indicates the first A set of core words for a priori topic information; This indicates that, except for the first element, the set of prior subject information contains... The weight coefficients of the core words of all prior topic information other than the first prior topic information.

3. The multi-dimensional topic discovery method based on prior information and boundary awareness according to claim 1, characterized in that, Based on the set of text units, the topic support level of each prior topic information in the prior topic information set is determined, specifically including: The core word coverage of each prior topic information is calculated using the following formulas: ； The expanded word coverage of each prior topic information is calculated using the following formulas: ； The semantic evidence coverage of each prior topic information is calculated using the following formulas: ； The topic support for each prior topic information is calculated using the following formulas: ； in, Indicates the first The coverage of core keywords in prior topic information; Indicates the first One core keyword; Indicates the first A set of core words for a priori topic information; Indicates the first The core words in the target text corpus Word frequency in; Indicates the first Expansion word coverage of prior topic information; Indicates the first One extended term; Indicates the first A set of negative boundary words for prior topic information; Indicates the first The extended words in the target text corpus Word frequency in; Indicates the first The semantic evidence coverage of prior topic information; Indicates the first Each text unit; Represents a collection of text units; Indicates an indicator function; Indicates the first The text unit vector corresponding to each text unit With the A positive topic prototype of prior topic information Cosine similarity; This indicates a preset semantic similarity threshold; Indicates the first The topic support of each prior topic information; , and These represent the weighting coefficients for core word coverage, extended word coverage, and semantic evidence coverage, respectively. If the first If the topic support of the first prior topic information is less than the first preset support threshold, then the... The prior topic information has a topic support level of insufficient evidence. If the first If the topic support of the first prior topic information is greater than or equal to the first preset support threshold and less than the second preset support threshold, then the first... The prior topic information has a weak level of support for the evidence. If the first If the topic support of the first prior topic information is greater than or equal to the second preset support threshold and less than the third preset support threshold, then the first... The thematic support level of the prior thematic information is the general level of evidence. If the first If the topic support of the first prior topic information is greater than or equal to the preset threshold of the third support, then the first... The thematic support level of each prior topic information is the level of sufficient evidence.

4. The multi-dimensional topic discovery method based on prior information and boundary awareness according to claim 1, characterized in that, For each valid prior topic information in the prior topic information set, based on the corresponding positive topic prototype and negative boundary prototype, soft evidence text paragraph retrieval is performed on the text unit set to obtain a candidate text evidence paragraph set corresponding to each valid prior topic information in the prior topic information set, specifically including: The word-literal matching score between any text unit in the text unit set and each valid prior topic information is calculated using the following formulas: ； The semantic similarity score and negative boundary score between any text unit in the text unit set and each valid prior topic information are calculated using the following formulas: ； ； The comprehensive evidentiary relevance between any text unit in the text unit set and each valid prior topic information is calculated using the following formulas: ； The soft purity score of any text unit in the text unit set relative to each valid prior topic information is calculated using the following formulas: ； in, Indicates the first The text unit and the first A word literal matching score for each valid prior topic information; Indicates the first One core keyword; Indicates the first A set of core words for a valid prior topic information; Indicates the weight coefficient of the core words; Indicates the first The first core keyword and the first BM25 word literal matching strength of each text unit; Indicates the first One extended term; Indicates the first An expanded set of words containing valid prior topic information; Indicates the weight coefficient of the expanded words; Indicates the first The first extended word and the first BM25 word literal matching strength of each text unit; Indicates the first One negative boundary word; Indicates the first A set of negative boundary words containing valid prior topic information; The weight coefficient representing negative boundary words; Indicates the first The negative boundary word and the first BM25 word literal matching strength of each text unit; Indicates the first The text unit and the first The semantic similarity score of each valid prior topic information; Indicates the first The text unit vector corresponding to each text unit With the A positive topic prototype of valid prior topic information Cosine similarity; Indicates the first The text unit and the first Negative boundary scores for each valid prior topic information; Indicates the first The text unit vector corresponding to each text unit With the Negative boundary prototype of valid prior topic information Cosine similarity; Indicates the first The text unit and the first The overall evidentiary relevance of each valid prior topic information; , and These represent the weight coefficients of the word literal matching score, the semantic similarity score, and the negative boundary score, respectively. Indicates the first The text unit and the first The soft purity score of each valid prior topic information; Indicates the first The text unit and the first The overall evidentiary relevance of each valid prior topic information; Indicates the first Background topic score for each text unit; Text units whose overall evidence relevance exceeds a preset relevance threshold are selected as the first... Add the candidate text evidence paragraphs containing the first valid prior topic information to the first... A set of candidate textual evidence paragraphs containing valid prior topic information.

5. The multi-dimensional topic discovery method based on prior information and boundary awareness according to claim 4, characterized in that, For each valid prior topic information in the prior topic information set, all candidate text evidence paragraphs in the candidate text evidence paragraph set are filtered to obtain the topic discovery results for each valid prior topic information, specifically including: For each valid prior topic information in the prior topic information set, all candidate text evidence paragraphs in the candidate text evidence paragraph set are added to the topic text evidence paragraph set corresponding to each valid prior topic information if the overall evidence relevance exceeds a preset relevance threshold and the soft purity score exceeds a preset soft purity score threshold. Candidate text evidence paragraphs with an overall evidence relevance exceeding a preset relevance threshold and a soft purity score not exceeding a preset soft purity score threshold are added to the cross-topic text evidence paragraph set corresponding to each valid prior topic information. Candidate words are extracted from the set of topic text evidence paragraphs corresponding to each valid prior topic information to obtain the candidate word set corresponding to each valid prior topic information. Each candidate word set corresponding to a valid prior topic is sorted and filtered using boundary awareness to obtain the topic discovery results corresponding to each valid prior topic.

6. The multi-dimensional topic discovery method based on prior information and boundary awareness according to claim 5, characterized in that, Each valid prior topic information is subjected to boundary-aware sorting and filtering of the candidate word set, resulting in topic discovery results for each valid prior topic information, including: The weighted occurrence intensity of each candidate word in the candidate word set corresponding to each valid prior topic information in the corresponding set of topic text evidence paragraphs of the valid prior topic information is calculated using the following formulas: ； The topic discriminative scores of each candidate word in the candidate word set corresponding to each valid prior topic information are calculated using the following formulas: ； The evidence score of each candidate word in the candidate word set corresponding to each valid prior topic information is calculated using the following formulas: ； The semantic relevance, cross-topic marginal value, and boundary risk of each candidate word in the candidate word set corresponding to each valid prior topic information are calculated using the following formulas: ； ； ； The final ranking score of each candidate word in the candidate word set corresponding to each valid prior topic information is calculated using the following formulas: ； in, Indicates the first The candidate word set corresponding to the _th valid prior topic information is the _th The candidate word in the first The weighted occurrence intensity of the thematic text evidence paragraphs in a set of valid prior thematic information; Indicates the first A set of thematic textual evidence paragraphs containing valid prior thematic information; Indicates the first The text unit and the first The soft purity score of each valid prior topic information; Indicates the first The candidate word set corresponding to the _th valid prior topic information is the _th The candidate word in the first Word frequency in each text unit; Indicates the first The candidate word set corresponding to the _th valid prior topic information is the _th The candidate word is relative to the first The topic discrimination score of each valid prior topic information; Indicates the first The candidate word set corresponding to the _th valid prior topic information is the _th The candidate word and the first A collection of thematic text evidence paragraphs containing valid prior thematic information. BM25 literal matching strength; Indicates the smoothing coefficient; Indicates the first The candidate word set corresponding to the _th valid prior topic information is the _th The candidate word and the first A collection of thematic text evidence paragraphs containing valid prior thematic information. BM25 literal matching strength; Indicates the first The candidate word set corresponding to the _th valid prior topic information is the _th The candidate word is relative to the first Evidence score for each valid prior topic information; This represents a balance coefficient between the weighted intensity and thematic distinctiveness; Indicates the first The candidate word set corresponding to the _th valid prior topic information is the _th The candidate word and the first The semantic relevance of a valid prior topic information to a positive topic prototype; Indicates the first The candidate word set corresponding to the _th valid prior topic information is the _th Candidate word vector representation of candidate words With the A positive topic prototype of valid prior topic information Cosine similarity; Indicates the first The candidate word set corresponding to the _th valid prior topic information is the _th The candidate word and the first The cross-topic marginal value of a valid prior topic information positive topic prototype; Indicates the first The candidate word set corresponding to the _th valid prior topic information is the _th Candidate word vector representation of candidate words With the A positive topic prototype of valid prior topic information Cosine similarity; Indicates the first The candidate word set corresponding to the _th valid prior topic information is the _th The candidate word and the first Boundary risk of a valid prior topic information positive topic prototype; Indicates the first The topic distribution entropy of candidate words; Indicates the first The candidate word set corresponding to the _th valid prior topic information is the _th The candidate word is relative to the first The final ranking score of each valid prior topic information; , and These represent the weighting coefficients for semantic relevance, cross-topic marginal value, and boundary risk, respectively. Based on the final ranking score, topic words are selected from the candidate word set corresponding to each valid prior topic information using a preset selection method and added to the topic word set corresponding to each valid prior topic information, while simultaneously generating boundary risk warnings; the preset selection method includes sorting each candidate word in the candidate word set corresponding to each valid prior topic information from high to low according to the final ranking score, and selecting the first preset number of candidate words as topic words; or, selecting candidate words in the candidate word set corresponding to each valid prior topic information whose final ranking score exceeds a preset threshold as topic words; The topic discovery results for each valid prior topic information are the topic term set, boundary risk warning, topic support level, and topic text evidence paragraph corresponding to each valid prior topic information.

7. A multi-dimensional topic discovery system based on prior information and boundary awareness, characterized in that, The multi-dimensional topic discovery system based on prior information and boundary awareness is used to implement the multi-dimensional topic discovery method based on prior information and boundary awareness as described in any one of claims 1-6, wherein the multi-dimensional topic discovery system based on prior information and boundary awareness comprises: The data acquisition module is used to acquire the target text corpus set and the corresponding prior topic information set; each prior topic information set includes a core word set, an extended word set, and a negative boundary word set. The text segmentation module is used to segment the target text corpus to obtain a set of text units. The prototype building module is used to build the positive topic prototype and negative boundary prototype for each prior topic in the prior topic information set. The topic support level determination module is used to determine the topic support level of each prior topic information in the prior topic information set based on the set of text units; the topic support level includes sufficient evidence level, moderate evidence level, weak evidence level, and insufficient evidence level; The topic labeling module is used to label prior topic information with a topic support level of insufficient evidence as insufficient evidence prior topic information, and to label prior topic information with a topic support level of sufficient evidence, moderate evidence, and weak evidence as valid prior topic information. The soft evidence retrieval module is used to perform soft evidence text paragraph retrieval on the text unit set for each valid prior topic information in the prior topic information set, based on the corresponding positive topic prototype and negative boundary prototype, to obtain a candidate text evidence paragraph set corresponding to each valid prior topic information in the prior topic information set; the soft evidence text paragraph retrieval includes word literal matching retrieval, semantic similarity retrieval, and negative boundary constraint retrieval. The topic discovery result generation module is used to filter all candidate text evidence paragraphs in the candidate text evidence paragraph set corresponding to each valid prior topic information in the prior topic information set, and obtain the topic discovery result for each valid prior topic information; the topic discovery result includes topic support level, topic word set, topic text evidence paragraph, and boundary risk warning.

8. A computer device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the multidimensional topic discovery method based on prior information and boundary awareness as described in any one of claims 1-6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the multi-dimensional topic discovery method based on prior information and boundary awareness as described in any one of claims 1-6.

10. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the multi-dimensional topic discovery method based on prior information and boundary awareness as described in any one of claims 1-6.

Images