Enterprise culture hot word analysis method based on content traceability

Abstract

The application discloses a content-tracing-based enterprise culture hot word analysis method, and relates to the technical field of enterprise culture analysis.The method comprises the following steps: collecting culture-related original data from three dimensions, and constructing an enterprise-specific culture term library through preprocessing, density peak clustering, de-redundancy and matching with top-level culture elements; classifying enterprise internal texts to be analyzed according to departments and text types, automatically matching terms of corresponding categories in the enterprise-specific culture term library, and performing directional word segmentation based on preset rules; calculating the scene coverage of terms according to department scenarios and text type scenarios, calculating the culture correlation degree by using Jaccard similarity, and screening core hot words by combining the Apriori algorithm; locking target hot word origin texts based on the core hot words, counting the track data of the origin text transmission channels, identifying the transmission nodes and constructing the transmission track; and the method improves the accuracy of hot word analysis and provides support for targeted promotion of enterprise culture.

Description

Technical Field

[0001] This invention relates to the field of corporate culture analysis technology, specifically a method for analyzing corporate culture hot words based on content tracing. Background Technology

[0002] Currently, existing methods for analyzing corporate culture buzzwords have the following significant shortcomings in practical applications, making it difficult to support enterprises' needs for precise management of cultural implementation:

[0003] First, the terminology database construction process lacks objective technical support and mostly relies on manual screening. It neither achieves redundancy removal through precise clustering driven by semantic similarity, nor establishes a systematic matching mechanism with the top-level cultural elements of the enterprise, which easily leads to problems such as terminology redundancy and deviation from the core culture.

[0004] Secondly, the hot word selection criteria are too simplistic, relying solely on word frequency statistics without comprehensively considering the breadth of the terminology's contextual coverage and the depth of its cultural relevance. Furthermore, the synergistic relationships between terms are not verified through association rules, making it difficult for the selected hot words to truly reflect the core of the company's culture.

[0005] Third, the lack of ability to trace the spread of trending words makes it impossible to locate the original texts and key dissemination nodes of trending words, making it difficult for companies to develop targeted strategies for cultural promotion.

[0006] Therefore, there is an urgent need for a method for analyzing corporate culture hot words based on content tracing. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides a method for analyzing corporate culture hot words based on content tracing, which solves the problems of redundant terminology, inaccurate hot word selection, and lack of dissemination tracing in existing hot word analysis.

[0008] To achieve the above objectives, this invention provides the following technical solution: a method for analyzing corporate culture hot words based on content tracing, comprising:

[0009] Step 1: Collect culture-related raw data from three dimensions: official corporate culture documents, internal corporate texts, and employee culture awareness survey data. Preprocess the raw data to obtain a high-frequency term candidate set, and perform density peak clustering on it to build a corporate-specific cultural term library.

[0010] Step 2: Classify the internal texts to be analyzed by department and text type, automatically match the corresponding terms in the enterprise's proprietary cultural terminology library, and perform targeted word segmentation based on preset rules;

[0011] Step 3: Calculate the terminology scenario coverage based on departmental scenarios and text type scenarios, calculate the cultural relevance using Jaccard similarity, and select core hot words using the Apriori algorithm.

[0012] Step 4: Based on the core hot words, locate the source text of the target hot words, count the trajectory data of the propagation channels of the source text, identify the propagation nodes and construct the propagation trajectory.

[0013] As a further aspect of the present invention, the specific operation for preprocessing the raw data is as follows:

[0014] Using Python's nltk or jieba libraries, perform cleaning operations, including removing punctuation marks, deleting common stop words, and correcting text errors;

[0015] Phrases are extracted from the cleaned text using binary or trigram segmentation modes.

[0016] The TF-IDF algorithm is used to calculate the TF-IDF value of the extracted phrases, and they are sorted in descending order. The top Num phrases are taken as the high-frequency term candidate set, where Num is set according to the actual situation.

[0017] As a further aspect of the present invention, the specific steps for performing density peak clustering on the high-frequency term candidate set are as follows:

[0018] Each candidate term is converted into a word vector using a Word2Vec pre-trained model. Then, cosine similarity is used to calculate the semantic similarity between any two terms, generating an n×n similarity matrix, where n is the number of candidate terms.

[0019] Calculate the similarity of all term pairs and take 1.2 times the median similarity as the cutoff distance dc;

[0020] For each term i, count the number of other terms with a similarity ≥ dc to obtain the local density. And find local densities greater than The value with the lowest similarity to i among all terms is used to obtain the relative distance. ,choose and Terms that are all in the top 30% are used as cluster centers;

[0021] For terms that are not cluster centers, calculate their semantic similarity with all cluster centers and assign them to the cluster with the highest similarity, forming multiple term clusters;

[0022] For each term cluster, the semantic similarity between all terms within the cluster is calculated, and a redundancy threshold Re is set. If there are terms that exceed the redundancy threshold, the term cluster is marked as a redundant term cluster. For redundant term clusters, the term with the highest TF-IDF value is retained, and finally, a deredundant term cluster is obtained. Each cluster corresponds to a cultural subcategory.

[0023] As a further aspect of the present invention, the specific operation for removing redundant terms from redundant term clusters is as follows:

[0024] Extract core semantic anchor words for each top-level cultural element, that is, 1-2 words that best represent the essence of the element, and record them in the top-level element classification table. The method for obtaining the top-level cultural elements is as follows: after obtaining the TF-IDF value of each word in the official corporate culture document, sort them in descending order, and take the top 20% of words as the extracted top-level cultural elements.

[0025] Extract the core semantic anchor words of each term in the redundancy-removing term cluster, and compare them with the anchor words of the elements in the top-level element classification table. Only the top-level elements with consistent anchor words are retained as candidate elements of the term.

[0026] For each term in the deredundant term cluster and candidate element, the matching degree is calculated using a word-level edit distance algorithm;

[0027] For each term, select the maximum similarity value max(sim) between it and all candidate elements. If max(sim) ≥ Sim, then retain the term in the redundancy removal term cluster, where Sim is the similarity threshold.

[0028] If the similarity of all candidate elements is less than Sim, then the term is marked as a term to be reviewed, and the Corporate Culture Department, in conjunction with the heads of relevant business departments, will make a judgment.

[0029] If a term does indeed have an implicit semantic connection with a top-level element, then retain it;

[0030] If a term is not associated with any of the top-level elements, it is removed from the redundancy term cluster.

[0031] After removing redundant terms from all the redundant term clusters, a corporate-specific cultural terminology library is constructed.

[0032] As a further aspect of the present invention, the specific steps for calculating similarity using the word-level edit distance algorithm are as follows:

[0033] Use Jieba segmentation to break down terms and candidate elements into words;

[0034] The Levenshtein library in Python is used to calculate word-level edit distance, which is the minimum number of steps required to transform one list into another through insertion, deletion, and replacement operations.

[0035] Calculate the similarity sim = 1 - (edit distance / max(number of terms, number of candidate element terms)).

[0036] As a further aspect of the present invention, the preset rules specifically include:

[0037] Extract all terms from the corresponding cluster, calculate the character length of the longest term, and set that length as the initial matching window value;

[0038] Terms are sorted from longest to shortest character length, and terms of the same length are arranged in the order of their original records in the terminology database to form a term matching sequence.

[0039] Input the term matching sequence and initial matching window value into the word segmentation tool, and set the priority matching rule according to the sequence order: the tool first matches the long term at the beginning of the sequence in the text with the largest window value. If the match is successful, it is marked as a complete unit. If it is not matched, it tries the shorter terms in the sequence in turn until the first round of matching is completed.

[0040] As a further aspect of the present invention, the specific steps for performing targeted word segmentation based on preset rules are as follows:

[0041] Open the word segmentation tool and turn off the automatic new word recognition function in the settings;

[0042] Select the forced matching mode based on preset sequences, import the categorized text into the tool, and perform word segmentation.

[0043] Extract the core semantic anchor words of all terms in the corresponding cluster, and retrieve all word segments in the text that contain these core semantic anchor words and have a length in the range [N1, N2] characters, where N1 and N2 are the lower and upper limits of the core semantic anchor words;

[0044] The retrieved word segments are filtered by character overlap: if the overlap between the word segment and the original term is ≥60%, it is marked as a potential term variant and compiled into a table in the format of "original term name + variant phrase + text number + number of occurrences".

[0045] As a further aspect of the present invention, the Jaccard similarity algorithm is used to calculate the correlation between each term and the top-level cultural element:

[0046] Perform semantic preprocessing, including synonym replacement and stop word removal, and break down the preprocessed terms and top-level elements into core word units;

[0047] Calculate Jaccard similarity = (intersection size of term and core word unit of top-level element) / (union size of term and core word unit of top-level element).

[0048] Based on the relationship between top-level elements and subcategories, an objective correction coefficient is set: if the term belongs to a direct subcategory of the top-level element, the correction coefficient is 0.4; if the term belongs to an indirect subcategory of the top-level element, the correction coefficient is 0.2; if the term has no classification relationship with the top-level element, the correction coefficient is 0.

[0049] Final cultural relevance = Jaccard similarity + correction factor.

[0050] As a further aspect of the present invention, the step of selecting core hot words using the Apriori algorithm is as follows:

[0051] Treat each text in the text library as a "transaction" and the terms in the text as "items in the transaction", and construct a "transaction-item" matrix;

[0052] Support is calculated as (number of transactions including both terms) / (total number of transactions), and confidence is calculated as (number of transactions including both terms) / (number of transactions including the first term).

[0053] Select term pairs that meet the criteria of "support ≥ 0.3 and confidence ≥ 0.7";

[0054] If a term in a term pair satisfies both "scene coverage ≥ threshold" and "cultural relevance ≥ threshold", then it is marked as a core hot word.

[0055] The method for determining the scene coverage and cultural relevance thresholds is as follows: collect scene coverage data for all terms and calculate the 75th percentile, which is used as the scene coverage threshold; collect cultural relevance data for all terms and calculate the 70th percentile, which is used as the cultural relevance threshold.

[0056] As a further aspect of the present invention, the propagation trajectory specifically includes an initiating node, a core propagation node, secondary propagation nodes, and a receiving node; the initiating node represents the creator of the originating text, i.e., the first proposer of the target hot word; the core propagation node represents the personnel or departments that both access the originating text and actively propagate the hot word in subsequent texts; the secondary propagation node represents the personnel or departments that only cite or forward it, but do not further disseminate it; and the receiving node represents the personnel or departments that only access or view it, but do not participate in its propagation.

[0057] This invention provides a method for analyzing corporate culture hot words based on content tracing, which has the following advantages compared with existing technologies:

[0058] (1) This invention collects cultural-related raw data from multiple dimensions and combines density peak clustering to achieve accurate redundancy removal of terms, establishes an objective matching mechanism with top-level cultural elements, effectively solves the problems of single data dimension and strong subjectivity of terminology database in existing methods, and makes the constructed enterprise-specific cultural terminology database more in line with business scenarios and core culture.

[0059] (2) In the hot word screening stage, this invention sets dual indicator thresholds based on scene coverage and cultural relevance, and combines the Apriori algorithm to verify the co-occurrence relationship of term pairs, thus avoiding the defect of existing methods that rely on a single word frequency, causing hot words to deviate from the core.

[0060] (3) This invention solves the problem that existing methods cannot locate the origin of hot words and key disseminators by locking the origin text, statistically analyzing multi-channel dissemination data, identifying dissemination nodes and constructing trajectories. Attached Figure Description

[0061] Figure 1 This is a flowchart of the steps of the present invention;

[0062] Figure 2 This is a flowchart illustrating the steps of using DPC to remove redundant terms from a high-frequency term candidate set in this invention.

[0063] Figure 3 This is a flowchart illustrating the steps involved in calculating correlation using Jaccard in this invention. Detailed Implementation

[0064] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0065] Example 1

[0066] like Figure 1 This invention provides a method for analyzing corporate culture hot words based on content tracing, including:

[0067] Step 1: Collect raw data related to culture from multiple dimensions, preprocess the raw data, and build a corporate-specific cultural terminology database;

[0068] The multiple dimensions specifically include official corporate culture documents, internal corporate texts, and employee culture awareness survey data.

[0069] Official corporate culture documents can be collected from the past three years, including the "Corporate Culture Manual," "Annual Strategic Plan," and "Employee Code of Conduct." The "Values ​​Chapter," "Strategic Goal Description," and "Behavioral Norms Clause" should be extracted as key points.

[0070] Internal enterprise documents are collected by category according to "department + document type". The departments include core business departments such as sales, technology, customer service, and human resources. The document types include department weekly reports, meeting minutes, customer service work orders, employee training notes, etc.

[0071] Based on the survey data on employee cultural awareness, a questionnaire was designed, which listed preliminary cultural phrases for selection, such as "customer needs first" and "colleagues help each other". The survey subjects covered all departments and levels.

[0072] Convert three types of data into a unified text format, store it in the "Original Database of Cultural Terms" on the enterprise's internal server, and mark the data sources, such as "Official Documents - 2024 Strategic Plan" and "Employee Survey - Grassroots Employees in the Technical Department".

[0073] For the text in the original database of cultural terms, preprocessing and initial extraction of high-frequency terms are required. The specific operations are as follows:

[0074] Use the nltk library or jieba library in Python to perform cleaning operations: ① Remove punctuation marks (including Chinese and English punctuation, special symbols); ② Delete common stop words; ③ Correct text errors (such as correcting "Customer demand response" to "Customer demand response").

[0075] Adopt a binary or ternary word segmentation pattern to extract phrases from the cleaned text. For example, the sentence "Improve customer satisfaction through technological innovation" can be segmented into "Technological innovation", "Customer satisfaction", and "Improve customer satisfaction".

[0076] Use the TF-IDF algorithm to calculate the TF-IDF values of the extracted phrases, sort them in descending order, and take the first Num phrases as the candidate set of high-frequency terms, where Num is set according to the actual situation.

[0077] After obtaining the candidate set of high-frequency terms, use density peak clustering DPC to process it and eliminate redundant terms to construct an enterprise-specific cultural term library.

[0078] Step 2: Classify the texts to be analyzed within the enterprise according to departments and text types, such as "Technical Department - Project Meeting Minutes" and "Customer Service Department - Work Order Records". Automatically match the terms in the corresponding categories in the enterprise-specific cultural term library according to the classification results. The specific operations are as follows:

[0079] Use the os library in Python to read the file storage path and intercept the "Department Name" field in the path. If there is no clear department information in the path, the department label can be extracted by reading the document author information to ensure no omission.

[0080] Establish a text type-keyword mapping table, perform keyword retrieval on the first Q characters before the document body, and mark the corresponding type if the number of hits for a certain type of keyword is ≥ q. For tabular files, read the header fields to match the type, where q is the quantity threshold, and both Q and q need to be set according to the actual situation of the document.

[0081] For example, if the path of a tabular document is "Enterprise Internal Document / Customer Service Department / 202406 / Customer Complaint Handling.xlsx" and the header contains three keywords: work order number, customer demand, and return visit result, it will automatically match the "Customer Service Department - Customer Service Work Order Record" cluster in the enterprise-specific cultural term library.

[0082] Based on a matching database of enterprise-specific cultural terms, terminology matching rules are constructed:

[0083] Extract all terms from the corresponding cluster, such as "customer demand response closed loop" and "customer demand response";

[0084] Calculate the character length of the longest term and set that length as the initial matching window value;

[0085] Terms are sorted from longest to shortest character length, and terms of the same length are arranged in the order of their original records in the terminology database to form a term matching sequence.

[0086] Input the term matching sequence and initial matching window value into the word segmentation tool, and set the priority matching rule according to the sequence order: the tool first matches the long term at the beginning of the sequence in the text with the maximum window value. If the match is successful, it is marked as a complete unit. If it is not matched, it tries the shorter terms in the sequence in turn until the first round of matching is completed.

[0087] Targeted word segmentation is performed based on the above matching rules, specifically including:

[0088] Open the word segmentation tool and turn off the automatic new word recognition function in the settings to prevent the tool from splitting terms into new words that the system automatically recognizes. For example, prevent "customer demand response" from being split into "customer demand" and "response".

[0089] Select the forced matching mode based on preset sequences, and import the categorized text into the tool to perform word segmentation.

[0090] Extract the core semantic anchor words of all terms in the corresponding cluster, and retrieve all word segments in the text that contain these core semantic anchor words and have a length in the range [N1, N2] characters, such as "timely response to customer needs" and "rapid response to customer needs" containing "customer needs" and with a length range of [6, 10] characters.

[0091] The retrieved word segments are filtered by character overlap: if the overlap between the word segment and the original term is ≥60%, it is marked as a potential term variant and compiled into a table in the format of "original term name + variant phrase + text number + number of occurrences" for future use in updating the terminology database.

[0092] Step 3: For the processed internal enterprise text to be analyzed, calculate the coverage of each term according to two dimensions: department and text type, to obtain the scenario coverage. The specific steps are as follows:

[0093] Departmental scenario coverage: List the core business departments of the enterprise. For each term, count how many departments' valid texts it appears in. If the text is a cross-departmental joint text, such as the minutes of a joint meeting between the technology department and the sales department, then when the term appears in that text, it is counted in the coverage statistics of both departments.

[0094] Text type scenario coverage: Identify the core text types and count how many types of effective texts a term appears in. For example, if "customer demand response" appears in the sales department's weekly report and the customer service department's work order, then the text type coverage is 2.

[0095] Calculate scene coverage = number of scenes covered by departments + number of scenes covered by text types;

[0096] The top-level cultural elements of the enterprise's proprietary cultural terminology library are invoked, and the Jaccard similarity algorithm is used to calculate the correlation between each term and the top-level cultural elements.

[0097] Based on scene coverage and cultural relevance, a dual-indicator threshold is set. The specific steps are as follows:

[0098] Collect scene coverage data for all terms and calculate the 75th percentile, which will be used as the scene coverage threshold.

[0099] Collect cultural relevance data for all terms and calculate the 70th percentile, which will be used as the cultural relevance threshold.

[0100] Association rule mining algorithms, such as the Apriori algorithm, are used to analyze the co-occurrence relationships between terms:

[0101] Treat each text in the text library as a "transaction" and the terms in the text as "items in the transaction" to construct a "transaction-item" matrix. For example, text 1 contains {customer demand response, 24-hour closed loop}, and text 2 contains {customer demand response, customer complaint follow-up}.

[0102] Support is calculated as (number of transactions including both terms) / (total number of transactions), and confidence is calculated as (number of transactions including both terms) / (number of transactions including the first term).

[0103] Select term pairs that meet the criteria of "support ≥ 0.3 and confidence ≥ 0.7";

[0104] If a term simultaneously meets the criteria of "scene coverage ≥ threshold, cultural relevance ≥ threshold, and high frequency co-occurrence with at least one term", it will be marked as a core hot word.

[0105] Step 4: Based on the core hot words, locate the source text of the target hot words, count the trajectory data of the propagation channels of the source text, identify the propagation nodes and construct the propagation trajectory;

[0106] The specific steps for locking the origin text of hot words based on core hot words are as follows:

[0107] From the obtained core keywords, select target keywords and retrieve all texts containing the target keywords in the enterprise document management system, sorting them in ascending order by creation time;

[0108] Compare the sorting results, find the text that first appears the target hot keyword, and record its key information: text name, department that created it, creation time, and creator.

[0109] If there are multiple versions of the text, confirm the specific version in which the target hot word first appeared by checking the version history, and correct the origin time to the version update time to ensure the accuracy of the starting point;

[0110] We statistically analyze the trajectory data of the originating text's dissemination channels from two aspects: document sharing channels and text citation channels.

[0111] The document sharing channel extracts access records of the source text through the access logs of the document management system, filters the access data within one month after the origin time of hot words, and records: the accessing department, the visitor, the access time, and the access behavior.

[0112] Text citation channels: Search the company's internal text library to find new texts within one month of the origin date that "clearly cite the source text" or "contain the target hot words", and record: the name of the cited text, the citing department, and the cited paragraph;

[0113] Based on the trajectory data obtained above, all personnel / departments related to the spread of trending words are identified, and the spread nodes are determined and the spread trajectory is constructed.

[0114] The propagation trajectory includes the initiating node, core propagation node, secondary propagation node, and receiving node;

[0115] The initiating node refers to the creator of the originating text, that is, the first person to propose the target hot word;

[0116] The core propagation node refers to the person or department that both accesses the origin text and actively propagates the hot words in subsequent texts;

[0117] The secondary propagation nodes refer to individuals or departments that only reference or forward the information but do not further disseminate it.

[0118] The receiving node refers to a person or department that only accesses or views the content but does not participate in its dissemination.

[0119] Starting with the original text, arrows connect each dissemination node, and the dissemination time and method are marked to construct a specific dissemination trajectory, providing direction for corporate culture promotion.

[0120] Example 2

[0121] This embodiment, based on Embodiment 1, further discloses a method for processing high-frequency term candidate sets and removing redundant terms using density peak clustering (DPC), such as... Figure 2 As shown, the specific content includes:

[0122] First, each candidate term is converted into a word vector using a Word2Vec pre-trained model. Then, cosine similarity is used to calculate the semantic similarity between any two terms, generating an n×n similarity matrix, where n is the number of candidate terms.

[0123] Calculate the similarity of all term pairs and take 1.2 times the median similarity as the cutoff distance dc;

[0124] For each term i, count the number of other terms with a similarity ≥ dc to obtain the local density. , The larger the value, the denser the semantic surroundings of the term;

[0125] For each term i, find a local density greater than The value with the lowest similarity to i among all terms is used to obtain the relative distance. , The larger the value, the more likely the term is to be a cluster center;

[0126] choose and Terms that are all in the top 30% are used as cluster centers;

[0127] For terms that are not cluster centers, calculate their semantic similarity with all cluster centers and assign them to the cluster with the highest similarity, forming multiple term clusters;

[0128] For each term cluster, the semantic similarity between all terms within the cluster is calculated, and a redundancy threshold Re is set. If there are terms that exceed the redundancy threshold, the term cluster is marked as a redundant term cluster. For redundant term clusters, the term with the highest TF-IDF value is retained, because a high TF-IDF value indicates that the term is more important in the document. Finally, the deredundant term clusters are obtained, and each cluster corresponds to a cultural subcategory, such as "customer service", "technological innovation", and "employee collaboration".

[0129] Extracting top-level cultural elements from corporate official cultural documents: After obtaining the TF-IDF value of each word in the corporate official cultural documents, sort them in descending order and take the top 20% of words as the extracted top-level cultural elements;

[0130] The semantic similarity between each term in the deredundant term cluster and the top-level cultural element is calculated as follows:

[0131] Extract core semantic anchor words for each top-level cultural element, that is, 1-2 words that best represent the essence of the element. For example, the anchor word for "customer first" is "customer", the anchor word for "innovation-driven" is "innovation", and the anchor word for "quality first" is "quality". Record them in the top-level element classification table as semantic guidance for initial matching.

[0132] Extract the core semantic anchor words of each term in the redundancy-removing term cluster, and compare them with the anchor words of the elements in the top-level element classification table. Only the top-level elements with consistent anchor words are retained as candidate elements of the term, and elements with irrelevant anchor words are excluded.

[0133] For each term in the deredundant term cluster and candidate element, the similarity is calculated using a word-level edit distance algorithm:

[0134] The term and candidate elements are split into words using Jieba segmentation. For example, the term "customer demand response" is split into [customer, demand, response], and the candidate element "customer first" is split into [customer, first].

[0135] The Levenshtein library in Python is used to calculate word-level edit distance, which is the minimum number of steps required to transform one list into another through insertion, deletion, and replacement operations.

[0136] Calculate similarity sim = 1 - (edit distance / max(number of terms, number of candidate element terms));

[0137] For each term, select the maximum similarity value max(sim) between it and all candidate elements. If max(sim) ≥ Sim, then retain the term in the redundancy removal term cluster. Here, Sim is the similarity threshold, which needs to be set according to the actual situation.

[0138] If the similarity of all candidate elements is less than Sim, then the term is marked as a term to be reviewed, and the Corporate Culture Department, in conjunction with the heads of relevant business departments, will make a judgment.

[0139] If a term does indeed have an implicit semantic connection with a top-level element, then retain it;

[0140] If a term is not associated with any of the top-level elements, it is removed from the deduplicated term cluster.

[0141] Example 3

[0142] This embodiment, based on Embodiments 1 and 2, further discloses a method for calculating the correlation between each term and the top-level cultural element using the Jaccard similarity algorithm, such as... Figure 3 As shown, the specific process is as follows:

[0143] Semantic preprocessing is performed as follows: ① Synonym replacement: Based on the synonym list in the enterprise's proprietary cultural terminology library, such as "response" = "feedback", "implementation" = "implementation", the terms are unified with the synonyms in the top-level cultural elements; ② Stop word removal: Meaningless auxiliary words in the terms and top-level elements are deleted, such as "and" and "and". After removing "and" from "cross-position and collaboration", it becomes "cross-position collaboration", ensuring that only core semantic characters are retained;

[0144] The preprocessed terms and top-level elements are broken down into core word units. For example, the term "customer demand response" is broken down into {customer, demand response}, and the top-level element "customer first" is broken down into {customer, first}.

[0145] Calculate Jaccard similarity = (intersection size of term and core word unit of top-level element) / (union size of term and core word unit of top-level element). For the example above: intersection = {customer}, size = 1, union = {customer, demand response, supreme}, size = 3, then Jaccard similarity = 1 / 3.

[0146] Based on the relationship between top-level elements and subcategories, set an objective correction coefficient:

[0147] If a term belongs to a direct subcategory of a top-level element, such as "Customer Needs Response" belonging to the direct subcategory "Customer Service" of "Customer First", then the correction factor is 0.4.

[0148] If a term belongs to an indirect subcategory of a top-level element, such as "cross-functional collaboration" belonging to the indirect subcategory "team support" of "employee win-win", then the correction factor is 0.2.

[0149] If the terminology has no classification association with the top-level element, then the correction factor = 0;

[0150] Final cultural relevance = Jaccard similarity + correction factor;

[0151] For example, the term "cross-functional experience sharing" is preprocessed into {cross-functional, experience sharing}, and the top-level element "employee win-win" is further subdivided into {employee, win-win}, resulting in a Jaccard similarity of 0 (no intersection). However, "cross-functional experience sharing" belongs to the indirect subcategory "team collaboration" of "employee win-win", resulting in a correction coefficient of 0.2. Therefore, the final cultural relevance is 0 + 0.2 = 0.2.

[0152] Some of the data in the above formulas are numerical calculations with dimensions removed, and the contents not described in detail in this specification are all prior art known to those skilled in the art.

[0153] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A method for analyzing corporate culture hot keywords based on content tracing, characterized in that, include: Step 1: Collect culture-related raw data from three dimensions: official corporate culture documents, internal corporate texts, and employee culture awareness survey data. Preprocess the raw data to obtain a high-frequency term candidate set, and perform density peak clustering on it to build a corporate-specific cultural term library. Step 2: Classify the internal texts to be analyzed by department and text type, automatically match the corresponding terms in the enterprise's proprietary cultural terminology library, and perform targeted word segmentation based on preset rules; Step 3: Calculate the terminology scenario coverage based on departmental scenarios and text type scenarios, calculate cultural relevance using Jaccard similarity, and select core hot words using the Apriori algorithm. The specific operation of calculating cultural relevance using Jaccard similarity is as follows: perform semantic preprocessing, including synonym replacement and stop word removal, and split the preprocessed terms and top-level elements into core word units. Calculate Jaccard similarity = (intersection size of term and core word unit of top-level element) / (union size of term and core word unit of top-level element); set an objective correction coefficient based on the association between top-level element and sub-category: if the term belongs to a direct sub-category of the top-level element, then the correction coefficient = 0.4; If a term belongs to an indirectly related subcategory of the top-level element, the correction coefficient is 0.2; if a term has no classification association with the top-level element, the correction coefficient is 0; the final cultural relevance is Jaccard similarity + correction coefficient; the specific operation of selecting core hot words using the Apriori algorithm is as follows: each text in the text library is regarded as a "transaction", and the terms in the text are regarded as "items in the transaction", constructing a "transaction-item" matrix; calculate support = (number of transactions containing two terms) / (total number of transactions), confidence = (number of transactions containing two terms) / (number of transactions containing the first term); select term pairs that satisfy "support ≥ 0.3 and confidence ≥ 0.7"; if a term in a term pair simultaneously satisfies "scene coverage ≥ threshold and cultural relevance ≥ threshold", then mark it as a core hot word; the method for obtaining the scene coverage and cultural relevance thresholds is as follows: collect scene coverage data for all terms and calculate the 75th percentile, using it as the scene coverage threshold; collect cultural relevance data for all terms and calculate the 70th percentile, using it as the cultural relevance threshold; Step 4: Based on the core hot words, locate the source text of the target hot words, count the trajectory data of the propagation channels of the source text, identify the propagation nodes and construct the propagation trajectory.

2. The method for analyzing corporate culture hot words based on content tracing according to claim 1, characterized in that, The specific steps for preprocessing the raw data are as follows: Using Python's nltk or jieba libraries, perform cleaning operations, including removing punctuation marks, deleting common stop words, and correcting text errors; Phrases are extracted from the cleaned text using binary or trigram segmentation modes. The TF-IDF algorithm is used to calculate the TF-IDF value of the extracted phrases, and they are sorted in descending order. The top Num phrases are taken as the high-frequency term candidate set, where Num is set according to the actual situation.

3. The method for analyzing corporate culture hot words based on content tracing according to claim 1, characterized in that, The specific steps for density peak clustering of the high-frequency term candidate set are as follows: Each candidate term is converted into a word vector using a Word2Vec pre-trained model. Then, cosine similarity is used to calculate the semantic similarity between any two terms, generating an n×n similarity matrix, where n is the number of candidate terms. Calculate the similarity of all term pairs and take 1.2 times the median similarity as the cutoff distance dc; For each term i, count the number of other terms with a similarity ≥ dc to obtain the local density. And find local densities greater than The value with the lowest similarity to i among all terms is used to obtain the relative distance. ,choose and Terms that are all in the top 30% are used as cluster centers; For terms that are not cluster centers, calculate their semantic similarity with all cluster centers and assign them to the cluster with the highest similarity, forming multiple term clusters; For each term cluster, the semantic similarity between all terms within the cluster is calculated, and a redundancy threshold Re is set. If there are terms that exceed the redundancy threshold, the term cluster is marked as a redundant term cluster. For redundant term clusters, the term with the highest TF-IDF value is retained, and finally, a deredundant term cluster is obtained. Each cluster corresponds to a cultural subcategory.

4. The method for analyzing corporate culture hot words based on content tracing according to claim 3, characterized in that, The specific steps to remove redundant terms from a cluster of redundant terms are as follows: Extract core semantic anchor words for each top-level cultural element, that is, 1-2 words that best represent the essence of the element, and record them in the top-level element classification table. The method for obtaining the top-level cultural elements is as follows: after obtaining the TF-IDF value of each word in the official corporate culture document, sort them in descending order, and take the top 20% of words as the extracted top-level cultural elements. Extract the core semantic anchor words of each term in the redundancy-removing term cluster, and compare them with the anchor words of the elements in the top-level element classification table. Only the top-level elements with consistent anchor words are retained as candidate elements of the term. For each term in the deredundant term cluster and candidate element, the matching degree is calculated using a word-level edit distance algorithm; For each term, select the maximum similarity value max(sim) between it and all candidate elements. If max(sim) ≥ Sim, then retain the term in the redundancy removal term cluster, where Sim is the similarity threshold. If the similarity of all candidate elements is less than Sim, then the term is marked as a term to be reviewed, and the head of the corporate culture department will make the judgment. If a term does indeed have an implicit semantic connection with a top-level element, then retain it; If a term is not associated with any of the top-level elements, it is removed from the redundancy term cluster. After removing redundant terms from all the redundant term clusters, a corporate-specific cultural terminology library is constructed.

5. The method for analyzing corporate culture hot words based on content tracing according to claim 4, characterized in that, The specific steps for calculating similarity using the word-level edit distance algorithm are as follows: Use Jieba segmentation to break down terms and candidate elements into words; The Levenshtein library in Python is used to calculate word-level edit distance, which is the minimum number of steps required to transform one list into another through insertion, deletion, and replacement operations. Calculate the similarity sim = 1 - (edit distance / max(number of terms, number of candidate element terms)).

6. The method for analyzing corporate culture hot words based on content tracing according to claim 1, characterized in that, The preset rules specifically include: Extract all terms from the corresponding cluster, calculate the character length of the longest term, and set that length as the initial matching window value; Terms are sorted from longest to shortest character length, and terms of the same length are arranged in the order of their original records in the terminology database to form a term matching sequence. Input the term matching sequence and initial matching window value into the word segmentation tool, and set the priority matching rule according to the sequence order: the tool first matches the long term at the beginning of the sequence in the text with the largest window value. If the match is successful, it is marked as a complete unit. If it is not matched, it tries the shorter terms in the sequence in turn until the first round of matching is completed.

7. The method for analyzing corporate culture hot words based on content tracing according to claim 1, characterized in that, The specific steps for performing targeted word segmentation based on preset rules are as follows: Open the word segmentation tool and turn off the automatic new word recognition function in the settings; Select the forced matching mode based on preset sequences, import the categorized text into the tool, and perform word segmentation. Extract the core semantic anchor words of all terms in the corresponding cluster, and retrieve all word segments in the text that contain these core semantic anchor words and have a length in the range [N1, N2] characters, where N1 and N2 are the lower and upper limits of the core semantic anchor words; The retrieved word segments are filtered by character overlap: if the overlap between the word segment and the original term is ≥60%, it is marked as a potential term variant and compiled into a table in the format of "original term name + variant phrase + text number + number of occurrences".

8. The method for analyzing corporate culture hot words based on content tracing according to claim 1, characterized in that, The propagation trajectory specifically includes initiating nodes, core propagation nodes, secondary propagation nodes, and receiving nodes. The initiating node refers to the creator of the originating text, i.e., the first person to propose the target hot word. The core propagation node refers to the person or department that both accesses the originating text and actively propagates the hot word in subsequent texts. The secondary propagation node refers to the person or department that only quotes or forwards the text but does not further disseminate it. The receiving node refers to the person or department that only accesses or reads the text but does not participate in its propagation.

Images