Service response method and device based on communication theme identification, equipment and medium

Abstract

The invention relates to the technical field of semantic recognition, can be applied to business scenes such as financial science and technology and medical health, and discloses a business response method, device and equipment based on communication theme recognition and a medium. And calculating the distribution density of the words in the plurality of text units, generating word weights in combination with word frequencies and the distribution density, identifying concerned topics, generating user feature tags or content classification identifiers, and associating with a service information system to trigger service operation. According to the method, the communication record is converted into the text data, the word frequency and the word distribution density are calculated, and the word weight with the discrimination degree is generated, so that a user concerned theme is automatically identified, and a label or a classification identifier is further constructed; automation of user behavior understanding, standardization of label generation and linkage of an operation process and a business system are realized, and personalized service response capability and operation efficiency are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of semantic recognition, and in particular to a business response method and device based on communication topic recognition, equipment and storage medium. BACKGROUND

[0002] In the field of financial technology business, social communication tools have become an important channel for customer relationship maintenance and market operation, especially when targeting high net worth customers, corporate customers and other sub-populations. The conversation content often contains a large amount of personalized service needs and behavior clues. However, the existing operation mode still mainly relies on batch message pushing and static customer tag management, making it difficult to capture potential financial needs expressed by users in communication in real time, such as attention to financial planning, loan process or risk products. Due to the strict data compliance requirements and information protection system faced by the financial industry, the data link between the communication platform and the internal business system is limited, making it difficult for operation personnel to obtain real behavior data of customers, and thus forming a dynamic customer portrait, and it is also difficult to carry out automatic operation response based on real-time behavior.

[0003] In the field of medical and health business, medical service providers increasingly communicate with patients through social communication platforms for daily communication and health management, such as pre-diagnosis consultation, rehabilitation follow-up, and chronic disease intervention. Text interaction is used to achieve information collection and service feedback. However, patient expression often has characteristics such as fragmentation, non-standardization and implicit preference, and existing manual methods cannot efficiently identify the health needs, symptom descriptions or risk tendencies implied in the conversation. This results in inaccurate identification of patient status and untimely response. In addition, due to the requirements of medical information security and compliance, unstructured interaction content in the front-end communication channel cannot be directly integrated into the internal information management system of the medical institution, limiting the landing of intelligent health management and personalized intervention capabilities.

[0004] In the field of intelligent text analysis technology, some methods can extract keywords and model topics from structured text such as web pages and documents, but there are still significant shortcomings in the face of natural language dialogue text in social platforms. On the one hand, existing methods generally ignore the distribution of words across multiple semantic fragments, resulting in inaccurate word weight calculation. On the other hand, there is a lack of mechanism for semantic mapping between high-weight keywords and pre-set topic sets, making it difficult to effectively identify the core focus of the conversation. In addition, the current label generation mechanism often stays in the static matching stage, lacking the ability to link dynamically generated labels or classification information with the background business system, making it difficult to achieve automatic label-driven business response or strategy triggering.

[0005] In summary, the prior art still has great defects in semantic recognition, theme classification, label generation and collaborative linkage with business systems of text interaction data, and cannot meet the depth requirements of user behavior understanding, the real-time requirements of service response and the rigorous requirements of data compliance in financial and medical scenarios. SUMMARY

[0006] The main purpose of the present application is to provide a business response method and device based on communication theme recognition, equipment and storage medium, aiming at solving the technical problems that the prior art cannot realize the automatic identification, theme classification and label generation of user behavior data in communication records, and efficiently link the generation results with business information systems to drive personalized operation.

[0007] To achieve the above purpose, the present application provides a business response method based on communication theme recognition, comprising:

[0008] Obtain the communication records in the communication platform, and convert the communication records into text data in a preset format;

[0009] Analyze the text data to determine the word frequency of the words in the text data;

[0010] Determine the cross-text unit distribution density of the words in the multiple text data units of the text data;

[0011] Determine the word weight of the words in the text data based on the word frequency and the cross-text unit distribution density;

[0012] Identify the focus theme of the user or content associated with the text data based on the word weight;

[0013] Generate a user feature label or content classification identifier according to the focus theme;

[0014] Associate the user feature label or content classification identifier with the business information system, and trigger the associated business operation.

[0015] Further, to achieve the above purpose, the present application provides a business response device based on communication theme recognition, comprising:

[0016] A communication data preprocessing module for obtaining communication records in a communication platform and converting the communication records into text data in a preset format;

[0017] A word frequency analysis module for analyzing the text data to determine the word frequency of the words in the text data;

[0018] A distribution density calculation module for determining the cross-text unit distribution density of the words in the multiple text data units of the text data;

[0019] A word weight calculation module, configured to determine the word weight of the word in the text data based on the word frequency and the cross-text unit distribution density;

[0020] a focus topic identification module, configured to identify focus topics of users or content associated with the text data based on the word weights;

[0021] A tag identification generating module, used to generate a user feature tag or a content classification identification according to the concerned topic;

[0022] The business linkage control module is used to associate the user feature tag or content classification identifier with the business information system and trigger the associated business operation.

[0023] Furthermore, to achieve the above-mentioned purpose, the present invention also provides a computer device, which includes a memory, a processor, and a business response program based on communication topic identification stored in the memory and runnable on the processor. When the business response program based on communication topic identification is executed by the processor, the steps of the business response method based on communication topic identification as described above are implemented.

[0024] Furthermore, to achieve the above-mentioned purpose, the present invention also provides a computer-readable storage medium, on which a business response program based on communication topic identification is stored. When the business response program based on communication topic identification is executed by a processor, the steps of the business response method based on communication topic identification as described above are implemented.

[0025] Beneficial effects: The present invention relates to the field of semantic recognition technology and can be applied to business scenarios such as financial technology and medical health. It discloses a business response method, device, equipment and medium based on communication topic identification, including: obtaining communication records in a communication platform and converting them into text data in a preset format, analyzing the text data to determine the word frequency, calculating the cross-text unit distribution density of words in multiple text data units, generating word weights based on the word frequency and distribution density, identifying the focus topics of users or content associated with the text data based on the word weights, generating user feature labels or content classification identifiers, and associating the labels or identifiers with business information systems to trigger corresponding business operations. The present invention converts communication records into structured text data and calculates the word frequency and distribution density to generate word weights with discrimination, thereby automatically identifying user focus topics and further constructing labels or classification identifiers to achieve automation of user behavior understanding, standardization of label generation, and linkage between operation processes and business systems, thereby improving personalized service response capabilities and operational efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] The present invention will be further described below with reference to the accompanying drawings and embodiments, in which:

[0027] Figure 1 A schematic diagram of an application environment of a service response method based on communication topic identification in an embodiment of the present invention;

[0028] Figure 2 This is a flow chart of an embodiment of a service response method based on communication topic identification according to the present invention;

[0029] Figure 3 A schematic diagram of functional modules of a preferred embodiment of a service response device based on communication topic identification according to the present invention;

[0030] Figure 4 A schematic diagram of the structure of a computer device according to an embodiment of the present invention;

[0031] Figure 5 FIG. 2 is another structural diagram of a computer device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0032] It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0033] The service response method based on communication topic identification provided by the embodiment of the present invention can be applied in the following situations: Figure 1 In an application environment, the user terminal communicates with the server terminal through a network. The server terminal can obtain the communication records in the communication platform through the user terminal and convert them into text data in a preset format, analyze the text data to determine the word frequency, calculate the cross-text unit distribution density of the word in multiple text data units, generate word weights based on the word frequency and distribution density, identify the user or content concerned topics associated with the text data based on the word weights, generate user feature tags or content classification identifiers, and associate the tags or identifiers with the business information system to trigger corresponding business operations. The present invention converts communication records into structured text data and calculates the word frequency and distribution density to generate word weights with discrimination, thereby automatically identifying user concerned topics, and further constructing tags or classification identifiers, thereby realizing the automation of user behavior understanding, the standardization of tag generation, and the linkage of operation processes and business systems, thereby improving personalized service response capabilities and operational efficiency. Among them, the user terminal can be, but is not limited to, various personal computers, laptops, smart phones, tablet computers and portable wearable devices. The server terminal can be implemented with an independent server or a server cluster consisting of multiple servers. The present invention is described in detail below through specific embodiments.

[0034] See also Figure 2 , Figure 2This is a flow chart of an embodiment of a service response method based on communication topic identification provided by the present invention. It should be noted that although a logical sequence is shown in the flow chart, in some cases, the steps shown or described may be performed in a different order than that shown here.

[0035] like Figure 2 As shown, the service response method based on communication topic identification proposed by the present invention includes the following steps:

[0036] S10, obtaining communication records in the communication platform and converting the communication records into text data in a preset format;

[0037] In this embodiment, to enable operational decisions based on user communication behavior, it is necessary to first obtain raw communication data. Acquiring communication records includes, but is not limited to, extracting message interaction data from instant messaging tools, customer service dialogue systems, online consultation modules, or messaging components embedded within the system. This data typically includes text, audio, images, files, timestamps, unique message identifiers, and sender and receiver identifiers. The data structure may be encapsulated in JSON, XML, or a vendor-specific proprietary protocol, and must be captured and synchronized through a communication interface or middle-layer data access module.

[0038] The original communication content must be uniformly converted into a processable text format. The text data conversion process involves parsing and transcribing multiple data formats. Text messages can directly read the text field; voice messages require the use of a voice recognition model to automatically transcribe speech to text; image messages containing text can use the OCR recognition module to extract the text; file content can call a structured parsing engine to extract text based on the file type (such as DOC, PDF). During the processing process, the associated indexes of the original information, such as timestamps, user IDs, and session IDs, must be retained to enable subsequent content tracing and user mapping.

[0039] The converted text should be formatted uniformly. This includes standardizing encoding (e.g., UTF-8), line breaks, and punctuation, removing system noise characters (e.g., emojis and control characters), and applying standard syntax to the text content (e.g., capitalization and punctuation). Field boundary specifications can also be set, such as clearly distinguishing between speaker identifiers, speech content, and structural tags, to form a standard field format, such as {user ID, session ID, message time, message content}.

[0040] Each sub-step in this stage needs to maintain context consistency at the semantic level and ensure that the information is traceable and structurally compatible in the subsequent processing chain. For example, the converted text data needs to be stored in a database or document system with field index structure to facilitate efficient access and processing by word frequency, distribution density, and other calculation steps.

[0041] To achieve communication record acquisition, various data access methods can be used, such as calling data access services provided by the communication platform through open API interfaces, or deploying data collection agents within the enterprise to listen to and cache communication data streams in real time. For voice messages, an ASR module based on deep neural networks is deployed for end-to-end speech recognition conversion, and the recognition results are filtered for confidence, and low-confidence segments are removed.

[0042] In the text analysis stage, a text cleaning module can be introduced to remove non-semantic symbols using regular expressions and language rule libraries, and to intelligently correct misspelled words. A template engine can be introduced in the formatting stage to define field output rules in a configuration manner, achieving consistent structure output for multi-platform and multi-source communication data.

[0043] According to industry compliance requirements, communication data desensitization rules can be set to encrypt or anonymize identity recognition information (such as name, phone number, and ID number) during the conversion process, ensuring that the data processing process complies with regulatory constraints.

[0044] Example: In a financial business scenario, a customer inquires about the loan process through a mobile app, and the speech content is sent in voice form. The system accesses the real-time API stream of the enterprise communication platform to obtain the voice content, and converts the voice to text through the deployed ASR model, such as: "Can I apply for a loan of how much amount with this income". The formatted text is written into the standard field database and bound with the user ID and timestamp, providing data support for subsequent demand recognition and risk assessment.

[0045] In the medical health scenario, users consult chronic disease medication plans through online consultation platforms, sending pictures and text containing medical terminology and drug names. The system extracts the text of the prescription information in the picture through the image recognition module, and combines the text message content such as "what should I replace after finishing this medicine", and processes it into a standard format text, and maps the text with the user's identity and conversation structure, supporting subsequent medical intent recognition and personalized recommendation modules for processing.

[0046] This embodiment, through the unified collection and formatting of communication records, can convert multi-source, multi-type user conversation data into a standardized text format, forming a unified, analyzable corpus data foundation. This mechanism effectively addresses the problem of behavioral mining being hindered by the diverse structure and unstructured content of raw data, improving the feasibility of subsequent semantic modeling and label generation, and providing high-quality input data for system-level user behavior identification and operational strategies.

[0047] S20, analyzing the text data to determine the frequency of words in the text data;

[0048] In this embodiment, after the unified format text data is constructed, it is necessary to perform content analysis to extract representative information in the language structure. The text data analysis includes three processes: word segmentation, word item statistics, and frequency normalization.

[0049] Word segmentation breaks down continuous natural language text into single or multiple meaningful language units. For Chinese corpora, high-precision word segmentation can be achieved using statistical probability models (such as Hidden Markov Models (HMMs)), rule-based dictionary matching methods, or deep learning-based word segmentation models (such as the BiLSTM-CRF architecture). For example, for the text "The customer wishes to apply for financial services," the word segmentation module breaks it down into "customer / desire / apply / financial management / business."

[0050] After completing the word segmentation operation, a word statistics table structure needs to be established. By traversing all the word segmentation units, the total number of times each word appears in the entire text data needs to be counted. This statistical operation is usually implemented based on a hash map structure, recording the cumulative number of occurrences of each unique word. This step aims to quantify the absolute frequency of each word in the corpus and characterize its importance in semantic communication.

[0051] To avoid bias caused by text length differences or uneven vocabulary distribution, the raw word frequencies must be normalized. This can be achieved using standard word frequency (TF, i.e., word frequency / total word count), logarithmic scaling, or a ranking normalization method based on word frequency ranking. The final output is the word frequency value for each individual word, forming a complete word frequency vector set that serves as the basis for subsequent topic extraction and feature modeling.

[0052] The entire processing process needs to maintain the consistency of the data structure of the three links: word segmentation, word frequency statistics and frequency normalization, to ensure that the frequency of each word can be traced back to the original text context.

[0053] In actual deployment, pre-trained Chinese word segmentation tools (such as Jieba, THULAC, and HanLP) can be used to build a word segmentation processing module, and dictionary expansion and custom entity recognition rules can be used to improve word segmentation accuracy. In particular, for texts where specific industry terms frequently appear, domain-specific terminology dictionaries can be configured, such as "net-worth financial management" and "credit rating" in the financial industry, or "hypertension" and "drug allergy" in the medical industry.

[0054] Word frequency statistics can be implemented using the MapReduce distributed computing framework, making them suitable for parallel processing of large-scale text data. During normalization, frequency thresholds can be set to filter out low-frequency terms, and a stop word list can be used to remove conjunctions and function words that don't contribute to semantic meaning, such as "的," "了," and "和."

[0055] The context window mechanism can also be combined to count the co-occurrence frequency of words to enhance the context-awareness of semantic analysis and provide multi-dimensional reference in subsequent word weight analysis and topic extraction.

[0056] Example: In a financial services scenario, a user repeatedly sends text messages such as "What are the returns on wealth management products?" and "Are there any recommendations for medium- and low-risk products?" After word segmentation, keywords such as "wealth management," "returns," "medium- and low-risk," and "recommendations" are extracted. A word frequency analysis reveals that the word "wealth management" appears frequently in multiple user corpora. After normalization, its word frequency value is significantly higher than other words, indicating that it has a high semantic weight in this type of user text, providing a basis for subsequent identification of user-focused topics.

[0057] In a healthcare scenario, a user enters the query "My child has had a recurring fever for two days and has taken ibuprofen but it hasn't gone down." The word segmentation system identifies the terms "child," "fever," "ibuprofen," and "fever reduction," and counts how frequently "fever" and "fever reduction" co-occur across multiple sessions. The normalized word frequency values ​​are used to infer that the user is highly interested in content related to children's fever reduction treatments, providing semantic support for subsequent recommendations for fever reduction guidelines or appointments with pediatricians.

[0058] By converting text data into standardized word frequency sets, this implementation quantifies the semantic significance of different terms in natural language, providing a high-quality feature foundation for subsequent keyword extraction and topic modeling. This mechanism effectively improves the ability to automatically identify user-focused content within text, facilitating automated and highly accurate user behavior analysis.

[0059] S30, determining the cross-text unit distribution density of the word in the plurality of text data units of the text data;

[0060] In this embodiment, to accurately identify the semantic impact of each word in the text data, it is necessary to evaluate the actual distribution of the word across the multiple text data units after division. A text data unit can be understood as the minimum structured processing granularity. Its division method is determined by the interaction boundaries or semantic blocks in the communication scenario, such as a customer service interaction, a valid reply, or a semantically complete record within a time window. Each unit is considered an independent information entity.

[0061] For each word, its appearance in different text units has a strong topic correlation, so it cannot be judged solely by its global word frequency. For this reason, a distribution density index across text units is introduced. Its calculation method is based on the distribution range of words in different units and normalized based on the total number of current text units. Furthermore, a logarithmic transformation mechanism is introduced to form an inverse document frequency (IDF) index. This index is defined as log(N / DF), where N represents the total number of text data units and DF represents the number of text data units in which the word actually appears.

[0062] If a word appears in most text units, its DF value is large, resulting in a smaller IDF value, indicating that it is a widely disseminated high-frequency word that may lack targeted semantic focus. Conversely, if a word appears in only a very small number of text units, its IDF value is high, reflecting its concentrated distribution and being more likely to be a representative keyword for a specific topic. During data processing, the logarithmic transformation has the effect of compressing long-tail distributions and reducing the weight of extremely high-frequency words, thereby avoiding misleading modeling due to the frequent appearance of certain generalized words.

[0063] To further enhance the stability of distribution density judgment, a Boolean matrix representation of word presence can be established for each text data unit, and DF parameters can be extracted based on this matrix. For original input text with incomplete semantics, chaotic structure, or blurred boundaries, a post-word segmentation sliding window strategy can be used to achieve robust unit division to ensure that the unit semantics have minimum self-consistency. In some languages, such as Chinese or in a mixed multilingual environment, preprocessing of word segmentation boundaries, part-of-speech tagging, and proper noun normalization is also required to reduce statistical errors and enhance the accuracy of semantic expression.

[0064] At the implementation level, TF-IDF can be used as a foundational metric model, using its classic term frequency-inverse document frequency structure to model distribution density. The system can use this weighted value as a feature input for subsequent model training or semantic clustering through an embedded vectorization module. In extended scenarios tailored to specific tasks, weighting factors can be introduced to adjust the weight calculation method for different types of terms in the IDF formula. For example, priority parameters can be introduced for industry terminology and user-defined keywords, further enhancing the system's recognition capabilities and feature sensitivity in professional contexts.

[0065] In one implementation, after converting communication records into text data, the system first divides the complete text into multiple text data units with semantic boundaries. The division granularity can be a single conversation, a user speech segment, or a context window. Then, word segmentation is performed on each text data unit to generate a corpus containing a word index.

[0066] For each word, the system counts whether it appears in each text data unit and obtains the document frequency DF corresponding to the word. Then, based on the total number of text data units N, the inverse document frequency formula in TF-IDF (IDF = log(N / DF)) is applied to generate the cross-text unit distribution density value of the word. To avoid instability caused by division by zero or near-zero, the system can introduce a smoothing term, such as using log(1+N / (1+DF)), to enhance the numerical stability of the model for low-frequency words.

[0067] In actual deployment, the system can integrate third-party open source text vectorization tools (such as scikit-learn's TfidfVectorizer) or build a customized lightweight computing module to cache the TF-IDF vector representation of each word in the feature mapping structure for subsequent use in identifying topics of interest and generating labels. For application scenarios that support multilingual input, the system can use adapted language models to process word segmentation and word frequency statistics in different languages, ensuring that distribution density calculations remain consistent in multilingual environments.

[0068] In another implementation, if the communication record exists in the form of images or voice, the system can first convert it into processable text through OCR or voice recognition modules, and then perform the above-mentioned text unit division and distribution density calculation operations in a unified manner. For transcriptions with low recognition accuracy, a confidence threshold can be introduced to filter out low-reliability terms to prevent misleading word density assessment results.

[0069] Example description: In the field of medical and health services, when patients communicate with intelligent customer service or human agents through online consultation platforms, the chat text often covers a variety of semantic blocks, such as description of the condition, historical medical history, medication consultation and examination items. After the system divides these conversation records into several text data units, it uses the TF-IDF model to calculate the cross-text unit distribution density of each medical-related word. For example, words such as "electrocardiogram", "preoperative evaluation" and "dyslipidemia" may only appear in one or two units, while common terms such as "hello", "excuse me" and "doctor" appear in multiple units. By calculating the IDF value, the system can identify those medical keywords with high density but low coverage and use them as candidate medical intent inputs to assist in generating diagnosis and treatment direction labels, thereby promoting the configuration of personalized health management strategies.

[0070] In the field of financial technology business, when customers consult through mobile terminals or service hotlines, their text information usually involves product questions, risk assessments, trading rules, and user demands. After the system divides the conversation content into several text data units, it calculates the distribution density of each financial keyword using the TF-IDF method. For example, in a customer's ten interactions, "fund stop-profit", "portfolio volatility", and "broad-based index" are only concentrated in a certain group of continuous units, while "financial management" and "income" are more widely distributed. With the help of this density, the system can determine that the latter is a background vocabulary, while the former has more personalized preference characteristics, and then mark the user's investment style, and promote product recommendations or risk warnings related to this style.

[0071] This embodiment calculates the inverse document frequency value in the TF-IDF structure. The system can not only measure the sparse distribution of words across different text units, but also effectively enhance the ability to identify rare but semantically significant keywords. The higher the density value, the more likely the word appears in a small number of units, and is more likely to be the core expression of a specific topic; while a low density value indicates that the word is spread across multiple units and may lack discriminability. By measuring the distribution density and introducing a logarithmic transformation to suppress the influence of extremely high-frequency words, the accuracy and robustness of subsequent topic identification, user modeling, and content label generation can be improved, thereby enhancing the system's ability to identify users' potential interests and contextual intentions.

[0072] S40, determining a word weight of the word in the text data based on the word frequency and the cross-text unit distribution density;

[0073] In this embodiment, in order to reflect the expression intensity and semantic differences of different words in the entire text data, it is necessary to integrate the word frequency information obtained in the previous step with the distribution density across text units to construct word weights for subsequent analysis and processing. The process first pairs the word frequency value of each word in the entire text with its corresponding distribution density value, and performs a numerical combination operation. A commonly used combination method is to perform a multiplication operation on the word frequency value and the distribution density value to generate a word weight value for each word. The weight value can indicate that the word is both high-frequency and concentrated in the text, or low-frequency but highly inclined to a specific topic.

[0074] Term frequency can be normalized to obtain a probability value, which is the ratio of the number of times a term appears in the entire text to the total number of terms. Distribution density uses the inverse document frequency (IDF) value after logarithmic transformation to suppress the influence of background terms. In specific implementations, the basic logic of the TF-IDF model can be used, using TF (Term Frequency) to represent the frequency of a term in a text and IDF (Inverse Document Frequency) to represent the rarity of a term across multiple text data units.

[0075] The resulting word weights form a word weight set, which can be used to screen high-impact keywords and guide subsequent topic identification and semantic classification. To improve stability and generalization, the word weight set can be normalized, such as by maximum normalization or Z-score normalization, to adapt it to different models or label generation strategies.

[0076] In practical implementation, a hash map or inverted index structure can be used to record the frequency and distribution density of each word, and vectorized operations can be performed to accelerate batch product operations. Alternatively, a sparse matrix structure can be used to represent the word frequency vector and the IDF vector, and matrix multiplication can be used to efficiently generate the word weight vector.

[0077] In some scenarios, a learnable weighting function can be introduced, taking word frequency and distribution density as inputs and using training to determine the optimal weight combination for different contexts. Words can also be filtered based on part of speech, retaining only nouns, verbs, and terminology to improve the discriminative power of the weighted results. Furthermore, in corpora with relatively uniform distribution density, an upper limit can be set for the IDF value to avoid over-amplifying the impact of unusually sparse words.

[0078] Example: In the healthcare sector, patients may repeatedly use high-frequency descriptive words like "pain" and "discomfort" when communicating with the platform. However, these words are widely distributed and lack specificity. By combining word frequency and distribution density, words that truly focus on a specific pathological topic (such as "interstitial lung disease" and "secondary prevention of coronary heart disease") are significantly more heavily weighted than other generalized words because they are concentrated in specific text units and have a moderate frequency, thereby improving the accuracy of disease diagnosis.

[0079] In the fintech sector, users may frequently use terms like "product," "return," and "recommendation" in their interactions with service providers, often with generalized meanings. By incorporating both frequency and density, terms like "FOF fund reallocation" and "credit spread expansion" receive higher weights due to their relatively low frequency but concentration in key units. This helps accurately identify users' investment priorities and drive the execution of personalized operational strategies.

[0080] This embodiment combines word frequency with distribution density across text units to effectively identify keywords that are semantically representative and have clustered distribution characteristics within the text. This avoids the problem of misjudging common words as high-weighted based solely on word frequency, and also avoids the problem of selecting low-impact words based solely on rarity. This strategy enhances the discriminability and robustness of semantic modeling, helping to improve the accuracy and stability of subsequent topic identification and tag generation.

[0081] S50, identifying a user or a topic of interest of the content associated with the text data based on the word weights;

[0082] In the present embodiment, in order to extract the content theme with semantic representativeness from the constructed word weight set, it is necessary to design a set of theme identification processes driven by weight. First, it is necessary to perform a screening operation on the word weight set, from which words that meet preset conditions are selected. These conditions can include that the weight value is greater than a specified threshold, the part of speech belongs to a noun or a compound term, or has been included in a domain vocabulary, etc. The words after screening are regarded as keywords, and these keywords have a stronger semantic guidance effect in the text data.

[0083] The keywords are then mapped to an existing set of topic categories. This mapping process can be achieved through vocabulary matching, embedding vector distance, and semantic similarity calculation. Each keyword may correspond to one or more predefined topic categories. To improve matching results, a weighting mechanism can be introduced during the mapping process, giving high-weighted keywords a greater contribution to the match.

[0084] After completing the initial mapping, count the frequency of each topic category mentioned in the mapping results. This frequency reflects the density of occurrence of that topic category in the current text data. By sorting all topic categories by frequency, select the category with the highest frequency as the dominant topic category. This dominant category represents the main semantic focus of the current text data and is considered the final topic of interest.

[0085] Once a topic is identified, it is written to the entity record of the user or content, driving subsequent tag generation, content push, or behavior modeling. This writing can be done using structured key-value storage to ensure compatibility with the structure of subsequent business systems.

[0086] In practical implementation, a Trie tree, hash map, or vector search engine can be used to quickly match keywords to topic categories, building a topic mapping index table to support high-concurrency calls. When filtering keywords, multiple weight threshold segments can be set, and different strategies can be applied to keywords within different segments. For example, extremely high-weighted keywords can be directly identified as significant topics, while medium-weighted keywords can be supplemented by contextual analysis.

[0087] The mapping method can use static vocabulary matching or introduce pre-trained semantic vector models (such as Word2Vec and BERT) to calculate the vector distance between keywords and topic category labels, thereby achieving more fine-grained topic classification. To enhance the adaptability of the model, a domain adaptation module can be introduced to dynamically expand the topic category vocabulary to incorporate emerging topics or industry terms.

[0088] When multiple candidate topics exist for the same text data, entropy weighting or weighted averaging can be used to combine and model multiple topic categories, thereby forming a topic distribution for multi-label classification. Keyword phrases that do not meet the coverage of the existing topic set can be marked as "unknown topic candidates" for manual review or subsequent vocabulary expansion.

[0089] Example: In the healthcare sector, patient text data may contain terms such as "chronic cough," "lung shadows," and "hormone dependence." After weighting, these terms are identified as high-weight keywords and mapped to topic categories such as "respiratory diseases" and "chronic disease management." Statistics show that "respiratory diseases" appears most frequently in the mapping, so this category is identified as the topic of interest in the current user's text. This topic can then be used to provide relevant examination and follow-up reminders, or even scientific information push notifications.

[0090] In the fintech sector, user text may contain keywords such as "yield drawdown," "credit rating change," and "fixed investment plan adjustment." By matching topic categories such as "fixed income strategy," "credit risk identification," and "robo-advisory model," the company ultimately identifies "fixed income strategy" as a topic of interest. This identification result can drive business systems to recommend appropriate product portfolios and deliver market analysis reports, improving the accuracy of customer service.

[0091] This embodiment, by constructing a topic identification mechanism driven by word weights, can effectively extract highly expressive semantic features from text data and map them into topic categories suitable for structured analysis. While ensuring topic accuracy, it avoids the noise and redundancy associated with over-reliance on full-text semantic modeling, improving the efficiency and stability of topic extraction. This approach establishes a clear semantic foundation for subsequent personalized tag generation, recommendation engine optimization, and risk warning.

[0092] S60, generating a user feature tag or content classification identifier based on the concerned topic;

[0093] In this embodiment, once a topic of interest is identified, it must be further converted into a data structure, namely, a user feature tag or content classification identifier, that can be used by the system for access, management, comparison, or linkage. This process first requires clarifying the object type corresponding to the topic of interest, that is, determining whether the current text data is associated with user interaction behavior or a content expression medium. This object type determination can be based on data source identification, communication session context, data storage path, or field classification in the business system.

[0094] If a user is identified, the system constructs structured text based on the semantic characteristics of the user's topic of interest, describing their preferences, areas of interest, or behavioral tendencies. This text should be concise, recognizable, and operationally feasible. For example, it can generate descriptive text such as "focused on risk prevention and control" or "interested in technological growth." This descriptive text is further converted into a tagged format, including fields such as tag identifier value, tag type, tag source, generation time, and tag level, to generate user feature tags.

[0095] If the current object is a content entity, it is necessary to construct a categorized text representation of the subject matter expressed in the content and convert it into a structured content classification identifier. This identifier includes not only the category name but also attributes such as the multi-level directory structure, category credibility, category version, and language style, which are used in subsequent content management, retrieval, or recommendation services. Both user tags and content identifiers must be persistently stored in a tag database and indexed to support rapid query, update, and policy response for both individuals and the entire entity.

[0096] You can use a tag generation template library to automatically assemble description templates based on the topic of interest. For example, when the topic is "asset allocation optimization," the system automatically generates a tag text such as "preferring diversified investment." When converting to a structured tag format, you can combine tag naming conventions (such as camel case and underline) and tag encoding rules (such as the module number + tag sequence number) to uniformly generate tag identification values. Tag levels can be scored based on word weight, topic frequency, or source credibility. The tag source field can record whether the tag was generated by keyword recognition, rule deduction, or model judgment.

[0097] When generating content classification identifiers, a hierarchical subject classification system can be constructed, mapping identified topics to standard content catalogs and determining their classification identifiers through the mapping path. The system can modify or enhance classification results by integrating with domain standards or industry knowledge bases to ensure that the identifiers are consistent with the existing content structure in the business system. To support multilingual or multi-scenario environments, the field naming and storage structure of tags and classification identifiers can be designed to support language extensions and semantic versioning.

[0098] Example: In the healthcare sector, if the system identifies a patient's high interest in content related to "chronic disease management," it can automatically generate user-specific tags such as "Chronic Disease Intervention Attention" and "Treatment Cycle Emphasis," and assign a high level to these tags, triggering the system to provide personalized follow-up reminders, long-term treatment record interfaces, or medication update notifications. Furthermore, if a patient's submitted content repeatedly mentions a particular disease, the content classification will be marked as "Respiratory Disease - Chronic Stage," supporting physicians in case classification and research analysis.

[0099] In the fintech sector, the system identifies frequent mentions of keywords like "return uncertainty" and "asset volatility" in customer conversations and categorizes them as "risk control tendency" topics, generating user-specific tags such as "conservative preference" and "volatility sensitivity." Articles related to these topics are automatically tagged as "asset allocation - conservative strategy" after being analyzed by the system. This allows the system to push content to customer groups with high tag matching, thereby improving content reach and customer response efficiency.

[0100] This embodiment converts topics of interest into structured tags and identifiers, not only achieving the migration of text semantics to executable attributes, but also providing a unified entry point for subsequent policy linkage, service orientation, and content governance. The standardization of the tag format enables the integration of semantic representations across different sources, scenarios, and objects, thereby enhancing data consistency and reusability. This approach avoids the ambiguity and ambiguity of unstructured semantics and improves the efficiency and stability of user profile construction, content classification, and business response.

[0101] S70: Associating the user feature tag or content classification identifier with a business information system, and triggering an associated business operation.

[0102] In this embodiment, after the user feature label or content classification identifier is constructed, it is necessary to establish a corresponding relationship between the structured data and the existing business information system to realize the triggering of linkage behavior. In this process, it is first necessary to complete the construction of the mapping and matching relationship between the label and the business system functional module. The mapping can be achieved through pre-configured label classification and business interface association rules, or it can be automatically completed through the label interface adaptation logic provided by the business system. The role of the mapping relationship is to clarify the business execution channel or system service type pointed to by different labels or classification identifiers, thereby providing a basis for subsequent execution paths.

[0103] Once the corresponding relationship is established, a command data structure containing label or classification identification information must be generated. In addition to the identification field, this structure should also include parameters such as the source time, object identity, priority attributes, and expected response type. This command is used to drive the execution of actions in the corresponding module of the business system, including content push, service notifications, risk control policy updates, customized product recommendations, and work order triggering.

[0104] Business information systems may include customer relationship management systems, content recommendation systems, product and service configuration systems, data collection systems, or risk control systems. By invoking internal interfaces, the system parses received structured instructions into specific action parameters and drives the corresponding functional modules to perform the actual operations. After the operation is executed, the system must provide feedback on the execution results, such as a success status code, failure reasons, or processing path logs. These feedback is recorded in interaction logs or user operation records to support subsequent behavior tracing and policy optimization.

[0105] In practical implementation, a pre-defined tag-to-interface mapping table can be used to manage the call relationships between user feature tags or content classification identifiers and business system functional modules. Upon detecting a newly generated tag, the system searches the mapping table for its target interface. For example, if the tag is "High Net Worth Clients Focus on Overseas Investment," the target interface is mapped to the "Overseas Asset Module Recommendation Interface." This mapping relationship can be configured using static rules or automatically evolved based on historical call paths using machine learning models.

[0106] The process of triggering a business operation can be achieved by constructing a standardized interface call request. For example, a JSON structure containing a tag identifier, user identity identifier, request type, and request parameters is constructed and sent to the target business system interface via HTTP or a message queue. The interface then parses and executes the specified task. For processes that require cross-system collaboration in some scenarios, the triggering process can also be encapsulated as a workflow task, with the workflow engine controlling the execution order and status synchronization across modules.

[0107] Feedback mechanisms can be implemented through callbacks, status polling, or event notifications, writing execution status to user behavior logs or tag usage logs for subsequent analysis and evaluation. To ensure information security and consistency, authentication, data signatures, and interface access rights management mechanisms can be introduced during data transmission.

[0108] Example: In the healthcare sector, when the system generates a label for a patient with the "Chronic Disease Recurrence Risk Concern" tag, it automatically links to the follow-up plan module within the medical services system and pushes a follow-up schedule and health assessment questionnaire based on the patient's medical history. The system then triggers the record to be stored in the electronic medical record and provides feedback to the patient. If the label is "High-Risk Medication Consultation," the system can link to the pharmacy services module, automatically triggering the delivery of professional pharmacist services to meet the patient's personalized medication consultation needs.

[0109] In the fintech business, upon identifying a customer with the "interest rate sensitive" tag, the system automatically establishes a call connection with the fixed-income product pool in the intelligent recommendation module, recommending financial products with low current volatility to the customer. The push notification and customer response status are recorded in the CRM. If the tag indicates "risk preference changes," the system links with the risk assessment system to update the customer's risk profile and adjust their investment portfolio strategy template to ensure continuity and relevance of customer service.

[0110] This embodiment forms a data-driven automated response mechanism by linking user feature labels or content classification identifiers with the business information system, effectively solving the problem that user behavior data is difficult to play an immediate value in the business system. The structured expression of labels or classification identifiers and interface standardization enable the system to respond to diverse user characteristics in a consistent manner, improving service accuracy and operational efficiency. At the same time, the recording mechanism of operation-triggered behavior provides a reliable basis for subsequent behavior feedback analysis, user profile calibration, and strategy optimization, thereby building a closed loop between data flow and business behavior.

[0111] The present invention relates to the field of semantic recognition technology and can be applied to business scenarios such as financial technology and medical health. A business response method, device, equipment and medium based on communication topic identification are disclosed, including: obtaining communication records in a communication platform and converting them into text data in a preset format, analyzing the text data to determine the frequency of words, calculating the cross-text unit distribution density of words in multiple text data units, generating word weights based on the word frequency and distribution density, identifying the focus topics of users or content associated with the text data based on the word weights, generating user feature labels or content classification identifiers, and associating the labels or identifiers with business information systems to trigger corresponding business operations. The present invention converts communication records into structured text data and calculates the frequency and distribution density of words to generate word weights with discrimination, thereby automatically identifying user focus topics and further constructing labels or classification identifiers, thereby realizing the automation of user behavior understanding, the standardization of label generation, and the linkage of operation processes and business systems, thereby improving personalized service response capabilities and operational efficiency.

[0112] In one embodiment, the above step S20 includes:

[0113] S201, performing word segmentation processing on the text data to obtain a word sequence;

[0114] S202, traversing the word sequence and counting the number of occurrences of each independent word in the text data;

[0115] S203, aggregating the number of occurrences of all independent words to generate the total number of words;

[0116] S204, divide the number of occurrences of each independent word by the total number of words to generate the word frequency of the word.

[0117] In this embodiment, in order to construct a semantic input for structured modeling, the original text data needs to be standardized, and the word frequency is used as a basic measure of the importance of the text internal word. The word frequency needs to be implemented based on a multi-level processing flow. In the initial stage of processing, the goal of word segmentation processing is to construct a set of word fragments with clear boundaries. This processing relies on language models, word segmentation dictionaries, or statistical rule sets, and can support term recognition and named entity disambiguation in different fields.

[0118] After completing the word segmentation, the overall expression of the text is converted into a linear word sequence, and the word mapping and counting operations are performed on this structure. The counting structure can include hash mapping, inverted index vector, or prefix tree structure, and has low time complexity of read and write ability. When traversing the word sequence, each unique word is assigned an independent counting unit. Each time the word is identified, the counting unit is added. Until the end of the traversal, the complete word frequency mapping table is formed.

[0119] After all word counting is completed, all counting fields in the mapping table are accumulated to generate the total amount of word statistics. This total amount is used as the denominator for the normalization calculation of the word frequency. By dividing each word count by the total count, a set of frequency ratios is formed. The ratio reflects the relative occurrence tendency of the word in the entire corpus, has additivity and comparability, and lays the foundation for subsequent cross-corpus feature alignment and relative weight modeling.

[0120] The word frequency result can be stored in sparse vector format, key-value pair structure, or matrix form, which is convenient for input into classification models, clustering models, or semantic reasoning engines. In the frequency statistics process, stop word filtering, low frequency pruning, or high frequency item truncation strategies can also be introduced to enhance the discrimination ability of the frequency distribution and control data redundancy.

[0121] This processing flow is highly coupled with subsequent density calculation, TF-IDF fusion, and topic recognition. The word frequency not only provides criteria for word selection, but also provides basic constraint conditions in the label reasoning and semantic classification process. Its accuracy and resolution directly affect the stability and generalization performance of the overall text semantic modeling.

[0122] This embodiment can effectively capture the main expression elements of the information content in the text by converting text data into word frequency distribution, providing a structured feature basis for subsequent calculations. This combination of word segmentation and word frequency extraction can systematically reveal user expression preferences or information focus points, thereby improving the ability to model the semantic features of the text. Compared with the manual keyword extraction method, this method has the characteristics of strong scalability and high adaptability, can adapt to different text sources and industry contexts, and realize the rapid quantitative expression of text content.

[0123] In one embodiment, the above step S30 includes:

[0124] S301, dividing the text data into a plurality of text data units;

[0125] S302, traversing the plurality of text data units, detecting whether each word appears in each text data unit;

[0126] S303, for each word, counting the number of text data units in which the word appears in the plurality of text data units as a text data unit coverage number;

[0127] S304: Divide the total number of text data units by the number of text data unit coverage of each word, and perform logarithmic transformation to generate the cross-text unit distribution density of each word.

[0128] In this embodiment, in order to characterize the distribution and diffusion of words in different contextual units, it is necessary to establish a mapping relationship based on the structured segmentation of text data and the distribution status of word items. First, the text data is divided into multiple text data units. The division basis may include paragraph boundaries, dialogue turns, sentence-level pause symbols, or fixed-length sliding windows. Different division rules can be adapted to the characteristics of structured text and unstructured text respectively. Each division unit is regarded as a minimum semantic block, which is used to construct a reference space for the cross-domain diffusion of word items.

[0129] After the partitioning is complete, all text data units are traversed to identify whether each word appears in each unit. This operation can be accomplished by creating a local vocabulary or a Boolean matrix. In this Boolean matrix, each row represents a word, and each column corresponds to a text data unit. A value of 1 indicates that the word exists in the unit, and a value of 0 indicates that it does not. By scanning the column dimension, we can quickly determine how many different text data units each word actually appears in.

[0130] For each word, the number of text data units detected as present is counted based on the Boolean matrix above. This is called the text data unit coverage. A larger value indicates a word with greater horizontal diffusion and universality, and generally weaker semantic differentiation. Conversely, if the word is concentrated in a small number of units, it may have higher semantic aggregation.

[0131] To obtain a comparability metric, the total number of text data units is divided by the word's text data unit coverage to obtain the inverse document frequency measure for that word. Because this ratio can be significantly inflated by extremely low coverage numbers, it is subsequently scaled using a logarithmic transformation, which calculates the natural logarithm or base-10 logarithm of the ratio. This logarithmic transformation reduces the impact of extreme values ​​while enhancing the ability to distinguish between medium- and low-frequency coverage words.

[0132] The resulting cross-text unit distribution density can be used as an indicator of the semantic concentration of a word, providing a semantic granularity foundation for subsequent word weight calculation and focus topic identification. This indicator can be structurally expressed as a sparse vector or combined with word frequency results to construct a feature matrix.

[0133] This embodiment effectively quantifies the semantic diffusion of words by dividing text data into multiple, clearly structured units and detecting the distribution of words within these units. Using an inverse coverage number plus a logarithmic transformation to generate the distribution density across text units not only addresses the information distortion caused by differences in the distribution breadth of different words within the corpus, but also improves the model's ability to distinguish between semantically dispersed and highly concentrated words, further enhancing the accuracy and stability of topic extraction and feature aggregation.

[0134] In one embodiment, the above step S40 includes:

[0135] S401, pairing the word frequency of each word with the cross-text unit distribution density of the same word;

[0136] S402, performing a multiplication process on the paired word frequency and the cross-text unit distribution density to generate a word weight value of the corresponding word;

[0137] S403: Integrate the word weight values ​​of all words to generate a word weight set.

[0138] In this example, to consider both the local frequency and cross-text diffusion characteristics of a word when characterizing its relative importance within the text data, the previously acquired word frequency is paired with its corresponding cross-text unit distribution density. The word frequency represents the relative frequency of the word within the overall text data, with a value between 0 and 1 and a normalized value. The cross-text unit distribution density measures the concentration of the word across different text data units and is obtained through a logarithmic transformation.

[0139] During the pairing process, the frequency of each word is mapped to its distribution density across text units, forming a structured pairing data table. This table can be expressed as key-value pairs or a vector dictionary. Subsequently, a numerical multiplication operation is performed on each pair of paired values ​​to obtain a composite index for the word, weighted by both semantic and frequency dimensions. This multiplication operation can be viewed as a weighted semantic emphasis strategy in a probability space, simultaneously highlighting high-frequency, highly aggregated keywords while suppressing semantic noise and sparsely frequent words.

[0140] After calculating the weights of all words, we aggregate them into a word weight set, expressed as a sparse vector, embedding matrix, or other form. This set serves as input for subsequent tasks such as topic identification and user modeling. This word weight set not only preserves the regularity of the original vocabulary in the corpus but also incorporates contextual distribution characteristics, resulting in strong discriminative power and scalability.

[0141] This embodiment multiplies word frequency by its distribution density across text units to construct a weighted metric that combines both word occurrence intensity and semantic distribution. This effectively avoids the limitations of a single frequency metric, which is susceptible to interference from common words, while also improving the accuracy of topic extraction, interest identification, and semantic classification. The constructed word weight set can express corpus content with higher information density, significantly enhancing the adaptability of downstream tasks and the personalized accuracy of output.

[0142] In one embodiment, the above step S50 includes:

[0143] S501, selecting words that meet preset weight conditions from the word weight set as key words;

[0144] S502, mapping the key words to predefined topic categories to generate a mapping result;

[0145] S503, counting the frequency of occurrence of each topic category in the mapping result, determining the topic category with the highest frequency of occurrence as the dominant topic category, and setting the dominant topic category as the focus topic;

[0146] S504: Write the concerned topic into the attribute record of the user or content entity.

[0147] In this embodiment, on the basis of the existing set of word weights, in order to identify the theme information of the user or content entity involved in the text data, a weight screening operation needs to be performed on each word in the weight set. The operation takes the preset weight condition as the screening basis, and the setting of the weight condition can be dynamically adjusted according to the distribution characteristics of the word weight, for example, using the mean weighted standard deviation of the word weight to construct a floating threshold, or truncating according to the percentile of the word weight. The purpose of the screening process is to retain words that have strong semantic expression in a specific context, have directionality or discrimination ability to the text structure, and exclude high-frequency but low-information contribution general words.

[0148] The set of high-weight words screened is subjected to a theme mapping process. The mapping process is based on a predefined theme set, which can be composed of a keyword theme index library constructed by humans, a semantic node structure of a knowledge graph, an industry knowledge label system, or a deep learning semantic embedding model. The establishment of the mapping relationship can adopt one of the following mechanisms: one is to realize explicit label docking by constructing one-to-one or one-to-many mapping rules between words and themes; two is to calculate the similarity of word and theme semantic vectors through a pre-trained model or word vector space, and select themes with similarity exceeding a threshold as candidates; three is to construct a semantic path scoring mechanism, combining context, word relationship and adjacency measurement in theme space, and performing multi-level semantic mapping.

[0149] When the mapping result is aggregated, the frequency of each theme category indicates how many different high-weight words are successfully classified into the theme, and the higher the frequency, the stronger the semantic association between the theme and the current text data. By sorting the frequencies of all theme categories, the dominant theme category can be identified, which is the theme category that is most representative in semantics and is pointed to by the most keywords in statistics. The dominant theme category is the final focus theme.

[0150] After the focus theme is generated, it needs to be written into the structured record of the user or content entity. This writing can take the form of structured label field extension, attribute field addition, JSON format nesting or graph database node annotation, etc., to ensure that the theme information can be queried, filtered, analyzed and fed back in subsequent systems. In the scenario of continuous input of multiple rounds of text data, the time evolution trend of the focus theme, the theme weight fluctuation, and the theme transfer path can be further used to construct a dynamic focus track and behavior change model. This operation enables the semantic attributes of the user or content entity to be updated in the system for a long time, forming an iteratively evolving data asset foundation.

[0151] This embodiment, through a word-weighted filtering and topic mapping mechanism, can extract the semantic topics most relevant to the user or content from redundant text content, avoiding the subjective bias and processing bottlenecks caused by relying on full-text matching or manual annotation. This step implements an efficient and automated process for identifying topics of interest, not only improving the accuracy of user modeling but also providing semantically informed label input for subsequent personalized services and business linkage, significantly enhancing the system's ability to understand and respond to user intent.

[0152] In one embodiment, the above step S60 includes:

[0153] S601, determining the object type associated with the topic of interest;

[0154] S602, when the object type is a user, generating a user description text based on the semantic content of the focus topic;

[0155] S603, converting the user description text into a user identification format, and generating a user feature tag containing the user identification format;

[0156] S604, when the object type is content, generating a content classification text based on the semantic content of the focus topic;

[0157] S605, converting the content classification text into a content identification format, and generating a content classification identification containing the content identification format;

[0158] S606: Store the user feature tag or content classification identifier in a tag database.

[0159] In this embodiment, after the focus topic is generated, in order to achieve structured expression for users or content, it is necessary to further determine the type of the associated attributes of the focus topic. The focus topic is usually expressed in the form of a subject vocabulary set composed of high-weight keywords, and the semantic content expressed has a clear attribution tendency, so the object type can be determined based on this. Object types include two categories: users and content. Among them, if the focus topic focuses on reflecting the user's behavioral characteristics, service preferences or intention expressions, it is determined to be a user object; if the topic mainly describes the information type, dissemination field or media attributes, it is determined to be a content object. The identification of object types can be based on the results of semantic analysis and predefined topic attribution rules, and a machine learning model can also be introduced to perform mapping training from topics to object types.

[0160] In actual system implementation, to determine the object type associated with a topic of interest, semantic analysis results can be combined with a topic attribution rule library, or a trained classification model can be used to identify the object type. Specifically, a vocabulary set of candidate topics of interest is first constructed using high-weighted keywords extracted using TF-IDF, such as "interest rate fluctuation sensitivity," "credit card limit management," and "robo-advisory risk adaptation." This vocabulary set is then input into a semantic matching module and compared against a predefined topic attribution rule library. The rule library labels behavioral, preference, and intent-oriented topics (such as "risk tolerance assessment," "wealth protection demands," and "medical cost sensitivity") as user-oriented, while topics representing content attributes (such as "bond market analysis report," "cardiovascular disease trends," and "medical device innovation trends") are labeled as content-oriented. If the vocabulary set contains word roots related to user behavior (such as "preference," "consultation," "intention," "need," and "use"), and the topic context is generated by an individual user, the object is determined to be the user.

[0161] In another implementation, a topic object classification model can be trained, with a large amount of historical annotated data as the training set. The model input is the word vector or BERT encoding representation of the topic of interest, and the output is the object category (user or content). For example, when the model recognizes that the topic of interest "Monitoring of side effects of hypertension drugs" is associated with patient behavior many times in the training set, its output is a user object. For another example, if the topic is "Annual medical big data trend analysis", the model output is a content object. This method can enhance the automation and robustness of judgment in actual deployment and is suitable for large-scale topic parsing tasks.

[0162] When the object type is determined to be a user, a user description text needs to be generated based on the semantic content of the topic of interest. The text should accurately depict the user's behavior pattern or interest orientation under the topic dimension. The user description text can be constructed by calling the semantic mapping library corresponding to the topic tag, or by constructing the sentence structure based on the template filling strategy. Subsequently, the user description text is converted into a unified user identification format to ensure that it can be normalized in the tag database. The user identification format can adopt a data tag structure oriented to business operations, such as a key-value pair format, a nested attribute structure, or an enumeration code type expression, so as to facilitate linkage calls with the business information system.

[0163] If the object type is content, the domain attributes and communication characteristics of the content expression need to be extracted from the focus topic to generate content classification text. Content classification text can be generated based on the content's attribution path in the domain knowledge graph, or it can be constructed by matching keywords with existing content classification systems. Furthermore, this content classification text needs to be converted into a unified content identification format, typically using standardized label encoding methods such as hierarchical content tags and vector-based topic encoding to support content retrieval, distribution, or recommendation.

[0164] Ultimately, whether generating user feature tags or content classification identifiers, they must be written to the tag database for persistent storage. The tag database supports structured storage and indexing mechanisms, ensuring rapid query and update of tags by the subsequent operating system. This storage process must ensure the uniformity of the data structure and the integrity of the index structure to support the stable operation of large-scale tag systems.

[0165] This embodiment realizes the mapping from unstructured text semantics to business-recognizable identifiers by converting the focus topic into structured user feature labels or content classification identifiers, effectively bridging the interface gap between semantic recognition and system applications. This method not only enhances the ability of semantic analysis results to be implemented in the business system, but also supports the flexible adaptation of differentiated processing strategies for user dimensions and content dimensions under the same analysis framework through the intelligent judgment mechanism of object types. Furthermore, with the help of a unified user identification format and content identification format, the label data has good system compatibility and scalability, and supports being used as a basic data resource for personalized service recommendations, content distribution strategy formulation, or user portrait modeling in business information systems. Finally, by writing structured tags into the label database, the archiving and call closed loop of label information is completed, providing a stable data foundation for subsequent high-frequency, multi-dimensional label queries and business triggering operations.

[0166] In one embodiment, the above step S70 includes:

[0167] S701, matching the user feature tag or content classification identifier with a preset service operation trigger policy set;

[0168] S702, when there is a matching business operation triggering policy, determining a target business information system corresponding to the business operation triggering policy;

[0169] S703: Locate the service processing interface corresponding to the user feature tag or content classification identifier in the target service information system;

[0170] S704, sending a tag data message containing the user feature tag or content classification identifier to the target business information system by calling the business processing interface;

[0171] S705: Receive the business operation feedback information returned by the target business information system and record it in the tag application log.

[0172] In this embodiment, to automatically integrate user feature tags or content classification identifiers with business processes, it is first necessary to match these tags with a set of preset business operation trigger policies. This policy set can be stored in a policy control module and describes the mapping relationship between specific tags and corresponding business operations in the form of rules. The matching operation can be completed using a search mechanism based on tag key fields or by calculating the similarity between the semantic category to which the tag belongs and the keywords in the policy rules. When a successful match is found, the corresponding business information system identifier is read from the rule to confirm the target business information system.

[0173] Next, it's necessary to locate a business processing interface compatible with the tag identifier within the target business information system. The definition of a business processing interface includes the service name, communication protocol, parameter format, and security verification fields, which can be obtained from the system registry or the interface specification configuration table. Interface matching operations typically determine the call path and field mapping scheme based on the interface adaptation field in the tag data structure.

[0174] After the interface is located, the module is called through the interface to construct a tagged data message. This message should include metadata fields such as the identification value of the user feature tag or content classification tag, the tag generation time, the tag type, and the semantic source. It is encoded and encapsulated according to the interface protocol. The tagged data message can be sent through mechanisms such as HTTP, gRPC, and MQ message queues to ensure the timeliness and security of the data during transmission.

[0175] Once the target business information system receives and processes the tag data, it returns corresponding business operation feedback information, which typically includes a processing status code, result description, processing timestamp, and subsequent processing suggestions. After receiving this feedback information, the tag processing module writes it to the tag application log in a structured format. The log record includes fields such as the tag identifier, sending destination, interface response, and trigger time, which are used for subsequent audit analysis, failure retry, and effect evaluation.

[0176] Example: In a fintech business scenario, a bank's private wealth management department uses this system to process text communication data between high-net-worth clients and their financial advisors. The system receives a long text transcript from a communication between the client and their relationship manager on the platform, covering topics such as "Can family trusts be deployed across borders?", "The interest rate differential between fixed-term and floating-rate products?", and "Real estate mortgages and overseas asset risk diversification." The system first imports this communication transcript and extracts standardized text data through the communication platform interface. It then performs word segmentation, removing redundant conjunctions and formatting noise to form a standardized sequence of financial semantic words. After counting the frequency of each word, the system calculates the total number of words and generates a standardized word frequency value. To account for the distribution of words across different semantic segments, the system segments the communication text into multiple text data units (e.g., client questions, advisor responses, etc.) based on time slices and speaker roles. For each key word, the system detects the number of units in which it appears, divides the total number of units by the number of units covered, and then performs a logarithmic transformation to determine the distribution density across text units. This value reflects whether the keyword is concentrated in a specific topic segment, thereby reflecting its potential semantic coreness. The system then multiplies the word frequency value by the distribution density value to form a weighted set of terms. The system then sorts the terms in descending order by weight, filtering out keywords such as "family trust," "overseas allocation," "fixed-term wealth management," "interest rate spread comparison," and "asset isolation." Using a topic mapping model, the system associates these keywords with predefined topic categories such as "cross-border asset management," "preference for return stability," and "risk protection needs." After counting the mapping frequencies, the system identifies "cross-border asset management" as the dominant topic in the current conversation and writes this topic into the customer's tag record. The system automatically identifies the current subject as a customer user and, based on this topic, generates a text tag, "Preference for stable cross-border asset allocation," which is further converted into the structured tag identifier, wealth_global_strategy_user_tag. After this tag is written to the tag database, it automatically triggers a tag update event in the customer profiling system. During the tag strategy matching process, this tag matches a push rule for "customer preference for cross-border assets." The system automatically connects to the "Private Wealth Product Recommendation Engine" in the target business system, calls its API, and transmits the tag. The recommendation engine indicates that the customer meets the recommended list of the current batch of newly issued "overseas bonds + insurance" combination products. The system then automatically arranges to send customized financial information to the customer, and the entire business closed loop is completed without human intervention.

[0177] Within a health management service platform, this system is used to analyze multiple conversations between chronic disease patients and their telehealth consultations. For example, a patient with chronic hypertension had multiple online conversations with their healthcare provider over the course of a week, raising questions such as "Has my heart rate slowed significantly after taking medication?", "Is daytime dizziness related to side effects of antihypertensive medication?", and "Can I switch medications to manage side effects?" The system captures all of the patient's conversation data from the platform's communication log interface, cleans the format, and converts it into standard text input. The system then applies a word segmentation strategy informed by medical knowledge, breaking down the text into words and identifying highly relevant medical terms such as "antihypertensive medication," "heart rate changes," "dizziness," "side effects," and "drug substitution." After counting word frequencies, the system then segments the text into multiple paragraphs for each conversation or question-and-answer session. The system calculates the number of units in which each keyword appears, generating a cross-unit distribution density. Combined with the word frequency, the system then calculates the word weight for each keyword. Analysis of these weighted sets reveals that keywords are concentrated in areas such as "medication adherence issues," "side effects," and "drug substitution recommendations." The system maps high-weight keywords to medical subject libraries such as "drug tolerance management", "medication behavior supervision", and "side effect risk monitoring", and identifies "drug tolerance management" as the dominant theme. This theme is written into the user record, and the system further determines it as a user entity, constructs a text tag "There are drug side effect adaptation issues", encodes it into a structured identifier medication_tolerance_user_tag and writes it into the tag library. Subsequently, the system matches the tag strategy and identifies that the tag is associated with the medication compliance module of the "Chronic Disease Management Platform", and transmits the tag identifier to the module through the API interface. After receiving this information, the platform automatically dispatches clinical pharmacists to carry out the "Drug Evaluation Intervention Plan", generates electronic intervention recommendations, feeds them back to the system and records them in the health management log, realizing automatic linkage of the entire process from semantic recognition to medical intervention tasks.

[0178] This embodiment achieves deep coupling between labels and business processes by dynamically matching user feature labels or content classification identifiers with a preset set of business operation trigger policies, and using standardized interfaces and data packets to drive business information systems to trigger operations. Under this mechanism, semantic recognition results can not only form structured labels, but also automatically drive the actions of related business systems based on label characteristics, thereby improving the business utilization rate and response speed of data analysis results. In addition, the complete recording capability of label application logs further supports full-process monitoring, problem backtracking, and data closure, effectively enhancing the stability, controllability, and traceability of the system.

[0179] In one embodiment, a service response device based on communication topic identification is provided, and the service response device based on communication topic identification corresponds one-to-one to the service response method based on communication topic identification in the above embodiment. Figure 3 , Figure 3 This is a functional module diagram of a preferred embodiment of a service response device based on communication topic identification according to the present invention. It includes a communication data preprocessing module 10, a word frequency analysis module 20, a distribution density calculation module 30, a word weight calculation module 40, a topic identification module 50, a tag generation module 60, and a service linkage control module 70. Each functional module is described in detail below:

[0180] The communication data preprocessing module 10 is used to obtain the communication records in the communication platform and convert the communication records into text data in a preset format;

[0181] A word frequency analysis module 20 is used to analyze the text data and determine the word frequency of words in the text data;

[0182] a distribution density calculation module 30, configured to determine a cross-text unit distribution density of the word in a plurality of text data units of the text data;

[0183] A word weight calculation module 40 is used to determine the word weight of the word in the text data based on the word frequency and the cross-text unit distribution density;

[0184] A focus topic identification module 50 is used to identify focus topics of users or content associated with the text data based on the word weights;

[0185] The tag identification generating module 60 is used to generate a user feature tag or a content classification identification according to the concerned topic;

[0186] The business linkage control module 70 is used to associate the user feature tag or content classification identifier with the business information system and trigger the associated business operation.

[0187] In one embodiment, the word frequency analysis module 20 is specifically configured to:

[0188] Performing word segmentation processing on the text data to obtain a word sequence;

[0189] Traversing the word sequence and counting the number of occurrences of each independent word in the text data;

[0190] Aggregate the occurrence counts of all independent words to generate the total word count;

[0191] The number of occurrences of each individual word is divided by the total number of words to generate the word frequency of the word.

[0192] In one embodiment, the distribution density calculation module 30 is specifically configured to:

[0193] dividing the text data into a plurality of text data units;

[0194] traversing the plurality of text data units, detecting whether each word appears in each text data unit;

[0195] For each word, counting the number of text data units in which the word appears in the plurality of text data units as a text data unit coverage number;

[0196] Dividing the total number of text data units by the text data unit coverage number of each word, and performing a logarithmic transformation process to generate a cross-text unit distribution density of each word.

[0197] In an embodiment, the word weight calculation module 40 is specifically configured to:

[0198] Pairing the word frequency of each word with the cross-text unit distribution density of the same word;

[0199] Performing multiplication processing on the paired word frequency and cross-text unit distribution density to generate a word weight value corresponding to the word;

[0200] Integrating the word weight values of all words to generate a word weight set.

[0201] In an embodiment, the focus topic identification module 50 is specifically configured to:

[0202] Filtering words that meet a preset weight condition from the word weight set as key words;

[0203] Mapping the key words to predefined topic categories to generate a mapping result;

[0204] Counting the occurrence frequency of each topic category in the mapping result, determining the topic category with the highest occurrence frequency as a dominant topic category, and taking the dominant topic category as a focus topic;

[0205] Writing the focus topic into the attribute record of the user or content entity.

[0206] In an embodiment, the label identification generation module 60 is specifically configured to:

[0207] Judging the object type associated with the focus topic;

[0208] When the object type is a user, generating a user description text based on the semantic content of the focus topic;

[0209] Converting the user description text into a user identification format to generate a user feature label containing the user identification format;

[0210] When the object type is content, generating a content classification text based on the semantic content of the focus topic;

[0211] Converting the content classification text into a content identification format to generate a content classification identification containing the content identification format;

[0212] The user feature tag or content classification identifier is stored in a tag database.

[0213] In one embodiment, the service linkage control module 70 is specifically configured to:

[0214] Matching the user feature tag or content classification identifier with a preset business operation trigger policy set;

[0215] When there is a matching business operation triggering policy, determining a target business information system corresponding to the business operation triggering policy;

[0216] Locating a business processing interface corresponding to the user feature tag or content classification identifier in the target business information system;

[0217] By calling the business processing interface, sending a label data message containing the user feature label or content classification identifier to the target business information system;

[0218] Receive the business operation feedback information returned by the target business information system and record it in the tag application log.

[0219] In one embodiment, a computer device is provided. The computer device may be a server, and its internal structure diagram may be as follows: Figure 4 As shown. The computer device includes a processor, a memory, a network interface and a database connected via a system bus. The processor of the computer device is used to provide determination and control capabilities. The memory of the computer device includes a non-volatile and / or volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external user terminal via a network connection. When the computer program is executed by the processor, it implements the functions or steps of the service side of a business response method based on communication topic identification.

[0220] In one embodiment, a computer device is provided. The computer device may be a user terminal, and its internal structure diagram may be as follows: Figure 5As shown. The computer device includes a processor, memory, network interface, display screen and input device connected via a system bus. The processor of the computer device is used to provide determination and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external server via a network connection. When the computer program is executed by the processor, it realizes the functions or steps on the user side of a service response method based on communication topic identification.

[0221] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the following steps are performed:

[0222] Acquiring communication records in a communication platform and converting the communication records into text data in a preset format;

[0223] Analyzing the text data to determine the frequency of words in the text data;

[0224] Determining a cross-text unit distribution density of the word in a plurality of text data units of the text data;

[0225] Determining a word weight of the word in the text data based on the word frequency and the cross-text unit distribution density;

[0226] Based on the word weights, identifying topics of interest of users or content associated with the text data;

[0227] Generating a user feature tag or content classification identifier based on the concerned topic;

[0228] The user feature tag or content classification identifier is associated with the business information system, and an associated business operation is triggered.

[0229] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the following steps are implemented:

[0230] Acquiring communication records in a communication platform and converting the communication records into text data in a preset format;

[0231] Analyzing the text data to determine the frequency of words in the text data;

[0232] Determining a cross-text unit distribution density of the word in a plurality of text data units of the text data;

[0233] Determining a word weight of the word in the text data based on the word frequency and the cross-text unit distribution density;

[0234] Based on the word weights, identifying topics of interest of users or content associated with the text data;

[0235] Generating a user feature tag or content classification identifier based on the concerned topic;

[0236] The user feature tag or content classification identifier is associated with the business information system, and an associated business operation is triggered.

[0237] It should be noted that the above functions or steps that can be implemented by the computer-readable storage medium or computer device can be found in the relevant descriptions of the server side and the user side in the aforementioned method embodiment. To avoid repetition, they will not be described one by one here.

[0238] Those skilled in the art will appreciate that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0239] Those skilled in the art will clearly understand that for the sake of convenience and brevity of description, only the division of the above-mentioned functional units and modules is used as an example. In actual applications, the above-mentioned functions can be distributed and completed by different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0240] It should be noted that if any software tools or components other than those of the Company appear in the embodiments of this application, they are merely for illustration and do not represent actual use. The above embodiments are intended only to illustrate the technical solutions of the present invention, not to limit them. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some of the technical features therein with equivalents. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included in the scope of protection of the present invention.

Claims

1. A service response method based on communication topic identification, characterized in that: The following steps are involved: Acquiring communication records in a communication platform and converting the communication records into text data in a preset format; Analyzing the text data to determine the frequency of words in the text data; Determining a cross-text unit distribution density of the word in a plurality of text data units of the text data; Determining a word weight of the word in the text data based on the word frequency and the cross-text unit distribution density; Based on the word weights, identifying topics of interest of users or content associated with the text data; Generating a user feature tag or content classification identifier based on the concerned topic; The user feature tag or content classification identifier is associated with the business information system, and an associated business operation is triggered.

2. The service response method based on communication topic identification according to claim 1, characterized in that: Analyzing the text data to determine the frequency of words in the text data includes: Performing word segmentation processing on the text data to obtain a word sequence; Traversing the word sequence and counting the number of occurrences of each independent word in the text data; Aggregate the occurrence counts of all independent words to generate the total word count; The number of occurrences of each individual word is divided by the total number of words to generate the word frequency of the word.

3. The service response method based on communication topic identification according to claim 1, characterized in that: Determining the cross-text unit distribution density of the word in the plurality of text data units of the text data includes: dividing the text data into a plurality of text data units; Traversing the plurality of text data units, detecting whether each word appears in each text data unit; For each word, counting the number of text data units in which the word appears in the plurality of text data units as the text data unit coverage number; The total number of text data units is divided by the number of text data unit coverage of each word, and a logarithmic transformation is performed to generate the cross-text unit distribution density of each word.

4. The service response method based on communication topic identification according to claim 1, characterized in that: Determining a word weight of the word in the text data based on the word frequency and the cross-text unit distribution density includes: Pair the word frequency of each word with the distribution density of the same word across text units; Multiply the paired word frequency and the cross-text unit distribution density to generate the word weight value of the corresponding word; The word weight values ​​of all words are integrated to generate a word weight set.

5. The service response method based on communication topic identification according to claim 1, characterized in that: Identifying, based on the word weights, topics of interest of users or content associated with the text data, including: Selecting words that meet preset weight conditions from the word weight set as key words; Mapping the key words to predefined subject categories to generate mapping results; Counting the frequency of occurrence of each topic category in the mapping result, determining the topic category with the highest frequency of occurrence as the dominant topic category, and taking the dominant topic category as the focus topic; The focus topic is written into the attribute record of the user or content entity.

6. The service response method based on communication topic identification according to claim 1, characterized in that: Generating a user feature tag or content classification identifier based on the topic of interest includes: Determining the type of object associated with the topic of interest; When the object type is a user, generating a user description text based on the semantic content of the focus topic; Converting the user description text into a user identification format, and generating a user feature tag containing the user identification format; When the object type is content, generating a content classification text based on the semantic content of the focus topic; Converting the content classification text into a content identification format to generate a content classification identification containing the content identification format; The user feature tag or content classification identifier is stored in a tag database.

7. The service response method based on communication topic identification according to claim 1, characterized in that: Associating the user feature tag or content classification identifier with the business information system and triggering associated business operations, including: Matching the user feature tag or content classification identifier with a preset business operation trigger policy set; When there is a matching business operation triggering policy, determining a target business information system corresponding to the business operation triggering policy; Locating a business processing interface corresponding to the user feature tag or content classification identifier in the target business information system; By calling the business processing interface, sending a label data message containing the user feature label or content classification identifier to the target business information system; Receive the business operation feedback information returned by the target business information system and record it in the tag application log.

8. A service response device based on communication topic identification, characterized in that: The service response device based on communication topic identification includes: A communication data preprocessing module, configured to obtain communication records in the communication platform and convert the communication records into text data in a preset format; A word frequency analysis module, configured to analyze the text data and determine the word frequency of words in the text data; a distribution density calculation module, configured to determine a cross-text unit distribution density of the word in a plurality of text data units of the text data; A word weight calculation module, configured to determine the word weight of the word in the text data based on the word frequency and the cross-text unit distribution density; a focus topic identification module, configured to identify focus topics of users or content associated with the text data based on the word weights; A tag identification generating module, used to generate a user feature tag or a content classification identification according to the concerned topic; The business linkage control module is used to associate the user feature tag or content classification identifier with the business information system and trigger the associated business operation.

9. A computer device, characterized in that: The computer device includes a memory, a processor, and a business response program based on communication topic identification stored in the memory and capable of running on the processor. When the business response program based on communication topic identification is executed by the processor, the steps of the business response method based on communication topic identification as described in any one of claims 1-7 are implemented.

10. A computer-readable storage medium, characterized in that The storage medium stores a service response program based on communication topic identification, which, when executed by a processor, implements the steps of the service response method based on communication topic identification according to any one of claims 1 to 7.

Images