A call content analysis method and device, electronic equipment and storage medium

Abstract

This application proposes a method, apparatus, electronic device, and storage medium for call content analysis. The method includes: obtaining a semantic feature vector corresponding to the call text between a customer service representative and a customer in a call scenario; obtaining the semantic similarity between the semantic feature vector and each tag vector in a preset tag library; obtaining multiple target tags corresponding to the semantic feature vector based on the semantic similarity; obtaining tag classification results for the multiple target tags based on the multiple target tags and multiple preset similarity thresholds, wherein the tag classification results include primary tags, secondary tags, and related tags; and obtaining an analysis report of the call text based on the tag classification results. The above technical solution improves the accuracy and precision of call content analysis by obtaining the importance classification of multiple target tags through the semantic similarity between the semantic feature vector of the call text and the tag vectors in the tag library, as well as multiple similarity thresholds.

Description

Technical Field

[0001] This application relates to the field of call content analysis technology, and in particular to a call content analysis method, apparatus, electronic device and storage medium. Background Technology

[0002] Currently, most call content analysis technologies use keyword matching or simple rule engines for analysis and tag extraction. Keyword matching can only recognize fixed words, and it is difficult to accurately identify content with similar meanings but different expressions. For example, keyword matching may miss the latter, such as "the product is not easy to use" and "the product experience is poor". Therefore, how to improve the accuracy of call content analysis is an urgent problem to be solved. Summary of the Invention

[0003] This application provides a method, apparatus, electronic device, and storage medium for analyzing call content, which can improve the accuracy and precision of call content analysis.

[0004] A first aspect of this application provides a method for analyzing call content, the method comprising:

[0005] Obtain the semantic feature vectors corresponding to the text of the call between the customer service representative and the customer in the call scenario;

[0006] Obtain the semantic similarity between the semantic feature vector and each tag vector in the preset tag library;

[0007] Based on the semantic similarity, multiple target labels corresponding to the semantic feature vector are obtained;

[0008] Based on the multiple target tags and multiple pre-set similarity thresholds, the tag classification results of the multiple target tags are obtained, and the tag classification results include primary tags, secondary tags and related tags;

[0009] An analysis report of the call text is obtained based on the tag classification results.

[0010] In some possible embodiments, obtaining the label classification results of the multiple target labels based on the multiple target labels and multiple pre-set similarity thresholds includes:

[0011] Based on the comparison between the semantic similarity of the target tags and the preset multiple similarity thresholds, the multiple target tags are classified into primary tags, secondary tags and related tags. The second similarity threshold corresponding to the primary tag is greater than the third similarity threshold corresponding to the secondary tag, and the second similarity threshold corresponding to the secondary tag is greater than the first similarity threshold corresponding to the related tag.

[0012] In some possible embodiments, obtaining multiple target labels corresponding to the semantic feature vector based on the semantic similarity includes:

[0013] Based on the similarity between the first tag and the second tag in the preset tag library, the tags in the preset tag library are clustered, where the first tag and the second tag are both any tags in the preset tag library;

[0014] Based on the clustering results and the preset tag similarity threshold, all tags in the preset tag library are divided into high-association clusters, medium-association clusters, and low-association clusters, and each tag contained in the same association cluster is semantically related.

[0015] The target association clusters corresponding to the labels whose semantic similarity is greater than the first similarity threshold are obtained as multiple target labels, and the target association clusters are at least one of the high association cluster, medium association cluster and low association cluster.

[0016] In some possible embodiments, the label classification results of the multiple target labels are obtained based on the multiple target labels and multiple pre-set similarity thresholds, including:

[0017] When the target association cluster includes both high association clusters and medium association clusters, the labels with semantic similarity greater than the second similarity threshold are selected from the high association clusters as the main labels;

[0018] From the highly correlated clusters and the moderately correlated clusters, select the tags whose semantic similarity is less than or equal to the second similarity threshold and greater than the first similarity threshold as secondary tags;

[0019] Other tags are selected from the highly correlated clusters as correlated tags, wherein the other tags are tags that do not include the primary tag and the secondary tag.

[0020] In some possible embodiments, obtaining the semantic feature vector corresponding to the call text includes:

[0021] The call text is mapped into a number vector according to a preset vocabulary;

[0022] The numerical vector corresponding to the call text is input into a pre-obtained neural network model to obtain the semantic feature vector corresponding to the call text, wherein the neural network model is obtained by training the call text and the training labels corresponding to the training call text.

[0023] In some possible embodiments, the neural network model is obtained in advance by methods including:

[0024] A neural network architecture based on Bidirectional Long Short-Term Memory (BiLSTM) is constructed, comprising an input layer, an embedding layer, a BiLSTM layer, and an output layer.

[0025] An attention mechanism is added to the neural network architecture. The attention mechanism layer is located after the BiLSTM layer and before the output layer. It is used to perform weighted aggregation of the temporal features extracted by the BiLSTM layer.

[0026] In some possible embodiments, obtaining the analysis report of the call text based on the tag classification results includes:

[0027] Based on the call text, multiple target tags corresponding to the call text, and the preset weights of the primary tags, secondary tags, and associated tags among the multiple target tags, a call summary corresponding to the call content is obtained;

[0028] Obtain tag statistics information for the multiple target tags, including the frequency of occurrence of each target tag and the time period distribution of each target tag;

[0029] In response to the user's selection of a preset output chart type, a visual statistical chart of the label statistics information is output.

[0030] An analysis report of the call content is obtained based on the call summary and the visual statistical charts, and the analysis report conforms to a preset report template.

[0031] A second aspect of this application provides a call content analysis apparatus, the apparatus comprising:

[0032] The acquisition module is used to obtain the semantic feature vectors corresponding to the call text between customer service and customer in a call scenario;

[0033] The processing module is used to obtain the semantic similarity between the semantic feature vector and each tag vector in the preset tag library; obtain multiple target tags corresponding to the semantic feature vector based on the semantic similarity; and obtain the tag classification results of the multiple target tags based on the multiple target tags and multiple preset similarity thresholds, wherein the tag classification results include primary tags, secondary tags and related tags.

[0034] The analysis module is used to obtain an analysis report of the call text based on the tag classification results.

[0035] A third aspect of this application provides an electronic device including a processor and a memory, the memory storing a computer program, the processor executing the computer program to implement the steps of any of the methods provided in the first aspect of this application.

[0036] A fourth aspect of this application provides a computer-readable storage medium storing computer instructions that, when executed by a processor, implement the steps of any of the methods provided in the first aspect of this application.

[0037] The technical solutions provided in this application have at least the following beneficial effects:

[0038] This application proposes a method, apparatus, electronic device, and storage medium for call content analysis. The method includes: obtaining a semantic feature vector corresponding to the call text between a customer service representative and a customer in a call scenario; obtaining the semantic similarity between the semantic feature vector and each tag vector in a preset tag library; obtaining multiple target tags corresponding to the semantic feature vector based on the semantic similarity; obtaining tag classification results for the multiple target tags based on the multiple target tags and multiple preset similarity thresholds, wherein the tag classification results include primary tags, secondary tags, and related tags; and obtaining an analysis report of the call text based on the tag classification results. The above technical solution obtains multiple corresponding target tags by using the semantic similarity between the semantic feature vector of the call text and the tag vectors in the tag library, and classifies these multiple corresponding target tags according to the magnitude of the similarity threshold. Tags with higher semantic similarity are more important for call content analysis, achieving the fusion of call text and different classification tags, thereby improving the accuracy and precision of call content analysis. Attached Figure Description

[0039] Figure 1 This is a schematic diagram illustrating an application scenario of a call content analysis method proposed in an embodiment of this application;

[0040] Figure 2 This is a flowchart illustrating a call content analysis method proposed in an embodiment of this application;

[0041] Figure 3 This is a flowchart illustrating another call content analysis method proposed in an embodiment of this application;

[0042] Figure 4 This is a flowchart illustrating another call content analysis method proposed in an embodiment of this application;

[0043] Figure 5 This is a flowchart illustrating another call content analysis method proposed in an embodiment of this application;

[0044] Figure 6 This is a flowchart illustrating another call content analysis method proposed in an embodiment of this application;

[0045] Figure 7This is a flowchart illustrating another call content analysis method proposed in an embodiment of this application;

[0046] Figure 8 This is a schematic diagram of the structure of a call content analysis device proposed in an embodiment of this application;

[0047] Figure 9 This is a schematic diagram of the structure of an electronic device proposed in an embodiment of this application. Detailed Implementation

[0048] The technical solutions in the embodiments of this application will now be described with reference to the accompanying drawings.

[0049] To facilitate a clear description of the technical solutions in the embodiments of this application, the terms "first" and "second" are used in the embodiments of this application to distinguish identical or similar items with essentially the same function and effect. For example, "first instruction" and "second instruction" are used to distinguish different user instructions and do not limit their order. Those skilled in the art will understand that the terms "first" and "second" do not limit the quantity or execution order, and the terms "first" and "second" are not necessarily different.

[0050] It should be noted that in the embodiments of this application, the words "exemplarily" or "for example" are used to indicate examples, illustrations, or explanations. Any embodiment or design scheme described as "exemplarily" or "for example" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design schemes. Specifically, the use of the words "exemplarily" or "for example" is intended to present the relevant concepts in a specific manner.

[0051] Furthermore, "at least one" refers to one or more, while "more than one" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can mean: A alone, A and B simultaneously, or B alone, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one of a, b, and c can mean: a, or b, or c, or a and b, or a and c, or b and c, or a, b, and c, where a, b, and c can be single or multiple.

[0052] It should be noted that, in the embodiments of this application, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0053] Currently, most call content analysis technologies use keyword matching or simple rule engines for analysis and tag extraction. Keyword matching can only identify fixed words, and it is difficult to accurately identify content with similar meanings but different expressions, such as "the product is not easy to use" and "the product experience is poor." Keyword matching may miss the latter. For example, traditional solutions treat call content and tags separately, such as only counting the number of times the tag "customer complaints" appears, but failing to associate it with specific semantics. Rule engines, on the other hand, require a large number of complex rules to be manually formulated. Not only are the rule maintenance and update costs high, but they are also difficult to adapt to changing business needs and language expressions, such as the peak consultation period during e-commerce promotions.

[0054] Current mainstream analytics technologies, such as keyword matching and rule engines, suffer from superficial semantic processing limitations. Keyword matching can only identify fixed words and cannot capture semantically equivalent expressions. For example, "the product is not easy to use" and "the product experience is poor" are semantically similar, but keyword matching may miss the latter due to word differences. Rule engines rely on manually defined complex logic, such as "complaint + delay" triggering tags, which limits the accuracy of call content analysis.

[0055] In view of this, embodiments of this application propose a method, apparatus, electronic device, and storage medium for call content analysis. The method includes: obtaining a semantic feature vector corresponding to the call text between a customer service representative and a customer in a call scenario; obtaining the semantic similarity between the semantic feature vector and each tag vector in a preset tag library; obtaining multiple target tags corresponding to the semantic feature vector based on the semantic similarity; obtaining tag classification results for the multiple target tags based on the multiple target tags and multiple preset similarity thresholds, wherein the tag classification results include primary tags, secondary tags, and related tags; and obtaining an analysis report of the call text based on the tag classification results. The above technical solution obtains multiple corresponding target tags by using the semantic similarity between the semantic feature vector of the call text and the tag vectors in the tag library, and classifies these multiple corresponding target tags according to the magnitude of the similarity threshold. Tags with higher semantic similarity are more important for call content analysis, achieving the fusion of call text and different classification tags, thereby improving the accuracy and precision of call content analysis.

[0056] The following is a detailed explanation of an application scenario example of an embodiment of this application. For example, as shown below... Figure 1 As shown, it includes: terminal device 101 and server 102, wherein:

[0057] For example, terminal device 101 can be an electronic device with calling capabilities, such as a mobile phone, computer, or smartwatch, used to obtain the text of a call. These devices capture the voice streams of both parties in a call through a built-in microphone array or an external audio interface, and transcribe the speech into structured text using a local speech recognition engine or cloud-based ASR service. For example, terminal devices are typically deployed on the user side or at enterprise customer service seats, adaptable to various scenarios such as mobile office and home customer service. Their portability and real-time advantages make the call text acquisition process more flexible and efficient, providing a high-quality data source for subsequent deep semantic analysis on the server side.

[0058] For example, when actually deploying a call content analysis system, the choice of server 102 needs to be flexibly matched according to the business scenario: for example, financial institutions with extremely high data privacy requirements usually use local physical servers to complete the semantic feature extraction and label classification of all call texts in the internal data center, ensuring that sensitive information does not leave the enterprise network; while e-commerce platforms targeting a massive number of users tend to use cloud servers such as Alibaba Cloud ECS or AWS EC2, leveraging elastic scaling capabilities to cope with the high-concurrency call analysis needs during promotional periods, while reducing hardware maintenance costs. For customer service hotlines that require real-time quality inspection, operators can deploy edge computing servers in the local data center to preprocess speech transcription and preliminary feature extraction, uploading only key features to the cloud, which reduces bandwidth consumption and achieves millisecond-level response; if complex deep learning model inference is involved, dedicated AI servers, such as NVIDIA DGX Station, can significantly improve the processing speed of BiLSTM+attention mechanism models with their multi-GPU parallel computing capabilities. The server is used for the analysis of the aforementioned call texts.

[0059] For example, the application scenario of this application can be the acquisition and analysis of call text on a terminal device. This application does not impose any limitations on this aspect.

[0060] After understanding an exemplary application scenario of the embodiments of this application, as follows: Figure 2 As shown below, we will describe in detail a call content analysis method proposed in the embodiments of this application. The method includes:

[0061] Step 201: Obtain the semantic feature vector corresponding to the call text between the customer service representative and the customer in the call scenario.

[0062] For example, the first step is to obtain the text of the call between the customer service representative and the customer. The text of the call is obtained as follows:

[0063] For example, in a call analysis scenario for employees of a logistics company, call recordings between customer service representatives and customers are acquired in real time. For example, user call details, call recordings, and transcribed text data can be obtained from a large network via operator call detail records (CDRs). Specifically, this involves interfacing with the operator's system and obtaining call data at preset intervals according to predefined operator system interface specifications. The call data includes: user call detail records, the storage path of the call recording file, and the transcribed text information corresponding to the user call detail records.

[0064] The call data is stored in a dual-channel recording mode. Voiceprint recognition technology or call rules are used to determine whether the speaker is a customer service representative or a customer, thereby obtaining the user role. This aims to gain insights into customer intentions and employee evaluations. The differentiation of call subjects allows for more accurate analysis of the call content.

[0065] For example, during the transcription of a call, a mature speech recognition API can be called to convert the call recording into text. Specifically, according to the API documentation of the selected speech recognition service, the recording file is prepared. The recording file usually needs to meet specific format requirements. A request is sent to the API to obtain the returned text transcription result, thereby obtaining the call text.

[0066] The initial transcribed text is then post-processed to correct homophone errors, punctuation errors, etc. For example, this processing involves using language models, rule matching, and other methods to further optimize the text returned by speech recognition. For instance, homophone errors are corrected by searching a thesaurus, and punctuation is adjusted according to grammatical rules.

[0067] Then, the call text is converted to a unified encoding format, such as UTF-8, and redundant whitespace characters are removed to standardize the text format and make it meet the requirements of subsequent processing.

[0068] After obtaining the call text, preprocessing operations can be performed on the call text to obtain the preprocessed call text between customer service and customer in the call scenario. The preprocessing operations of the call text include cleaning the obtained call detail records, transcribed text and other data. The cleaning steps include: removing duplicate records, invalid characters, garbled characters, etc., and appropriately filling or marking missing values.

[0069] Exemplarily, the preprocessing operation may include selecting appropriate natural language processing tools for texts in different languages, such as tools like the Natural Language Toolkit (NLTK) or the Language Technology Platform (LTP) to segment the call text according to word boundaries, facilitating subsequent language analysis.

[0070] Exemplarily, the preprocessing operation may also include, through a词性标注模型 (not clear in Chinese, should be something like "part-of-speech tagging model"), marking the corresponding part-of-speech for each word segmentation result, such as nouns, verbs, adjectives, etc., which helps to understand the semantic structure of the text.

[0071] Exemplarily, the preprocessing operation may also include, through named entity recognition technology, identifying entities such as person names, place names, and organization names in the text, and using a named entity recognition model to extract entity names with specific meanings from the call text, providing a basis for subsequent semantic analysis and information extraction.

[0072] Exemplarily, the preprocessing operation may also include removing stop words in the text, such as meaningless words like "的", "了", "在", etc. According to a pre-constructed stop word list, filter out words in the text that contribute less to semantic expression, reducing the data volume and highlighting key information.

[0073] Exemplarily, the preprocessed call text can be processed through a neural network model, such as the Bidirectional Encoder Representations from Transformers (BERT) or Bidirectional Long Short-Term Memory Networks (BiLSTM) based on Transformers, to obtain the semantic feature vector corresponding to the call text.

[0074] In an embodiment of the present application, exemplarily, an attention mechanism layer may also be added to the neural network architecture. The attention mechanism layer is after the BiLSTM layer and before the output layer, and is used to perform weighted aggregation on the temporal features extracted by the BiLSTM layer. By combining BiLSTM with the attention mechanism and adding an attention mechanism layer on the basis of the BiLSTM network, the model can automatically focus on important semantic parts when processing text, improving the model's ability to understand key information.

[0075] Step 202: Obtain the semantic similarity between the semantic feature vector and each label vector in the preset label library.

[0076] The preset tag library mentioned in this step can be dynamically added to or removed from, covering various business-related tags such as "product consultation," "customer complaints," and "service recommendations." For example, the tag library can be created in the form of a data structure, such as a dictionary or list, supporting the addition, deletion, or modification of tags at any time. Simultaneously, some metadata, such as tag descriptions, can be associated with each tag.

[0077] For example, the preset tag library may be obtained in advance based on experience or based on historical call texts, and this application embodiment does not limit it in this way.

[0078] For example, the preset tag library, obtained based on historical call text, includes:

[0079] Based on the matching degree between the historical call text and the tags in the preset tags, if the matching degree is less than the preset matching degree, a new tag corresponding to the call text is obtained, and the new tag is added to the preset tag library. In other words, if the matching degree is too low, the call text undergoes new analysis and feature extraction to generate a new tag, which is then added to the preset tag library.

[0080] For example, the preset tags include multiple tags, and each preset tag can be trained by a pre-trained word vector model or by random initialization to obtain a tag vector. The tag vector value is the vector of semantic features corresponding to each tag. Then, the cosine similarity between the semantic feature vector and these tag embedding vectors is calculated to measure the degree of association between the text and the tags.

[0081] For example, the cosine similarity value ranges from -1 to 1. The closer the similarity is to 1, the more similar they are; the closer it is to -1, the less similar they are; and a value close to 0 indicates that the correlation between the label vector and the semantic feature vector is very low.

[0082] For example, the method for obtaining label vectors can be vector similarity calculation based on Scikit-learn or vectorized calculation based on TensorFlow. Vectorized calculation based on TensorFlow is suitable for integration within the model, and the embodiments of this application do not limit the calculation method.

[0083] Step 203: Based on the semantic similarity, obtain multiple target labels corresponding to the semantic feature vector.

[0084] The semantic feature vector and the semantic similarity of the label vector obtained by the above method are used to obtain multiple target labels when the similarity threshold is greater than the first similarity threshold. The first similarity threshold can be a threshold with low relevance. Subsequently, these labels are classified among the relevant labels with a similarity greater than the first similarity threshold.

[0085] For example, the multiple target tags can be obtained directly from a preset tag library based on a first similarity threshold, or the tags in the preset tag library can be clustered into high-relevance clusters, medium-relevance clusters, and low-relevance clusters. Then, relevant clusters with a higher similarity than the first preset similarity can be obtained from the preset tag library, and the tags in the relevant clusters can be further classified.

[0086] Step 204: Based on the multiple target tags and multiple pre-set similarity thresholds, obtain the tag classification results of the multiple target tags. The tag classification results include primary tags, secondary tags, and related tags.

[0087] For example, primary tags refer to those tags that are highly important and unique. They typically directly reflect the core theme or key characteristics of the data and are the most crucial part of data classification and retrieval. Secondary tags are tags that are related to primary tags but are of relatively lower importance. They can supplement the information of primary tags and help to describe the data more comprehensively, but they do not play a decisive role like primary tags. Related tags are those tags that have some connection with primary and secondary tags but are less similar. They may belong to the same data domain or theme but differ in specific descriptions. Classifying them as related tags can better organize and present the diversity of data.

[0088] Compared to existing technologies, the embodiments of this application break through the limitations of traditional manual rules. Preset tags can be automatically expanded according to business needs and semantic mapping can be learned through model training, eliminating the need for manual rule maintenance. Multi-tag hierarchical association is achieved based on multiple similarity thresholds of 0.85 for the primary tag and 0.6 for the secondary tag, which is suitable for complex call scenarios. Existing customer profiling technologies rely on multi-dimensional rule integration and lack deep natural language fusion. However, the embodiments of this application achieve automated and accurate tag allocation through pure data-driven cosine similarity calculation, solving the efficiency bottleneck of traditional semi-supervised learning that requires manual rule adjustment.

[0089] Step 205: Obtain an analysis report of the call text based on the tag classification results.

[0090] For example, the step of obtaining an analysis report of the call text based on the tag classification results, such as... Figure 3 As shown, the method includes the following steps:

[0091] Step 301: Obtain a call summary corresponding to the call content based on the call text, multiple target tags corresponding to the call text, and the preset weights of the primary tags, secondary tags, and associated tags among the multiple target tags.

[0092] For example, the call text obtained after the above steps and the corresponding tag results are input into a Natural Language Generation (NLG) library such as TextGenrnn. The NLG library automatically generates a concise and accurate call content summary based on its pre-trained language model and understanding of the call text's semantics. The preset weight of the primary tag is greater than the preset weight of the secondary tag, and the preset weight of the secondary tag is greater than the preset weight of the associated tag. A clear hierarchy of call content summaries is generated based on the different weights of the tags. During this process, the model analyzes key information in the call text. This key information can be information indicated by the primary tags, such as customer feedback issues or requests. Combining this with the thematic direction represented by the tags, the core content is extracted. For example, if the call text revolves around customer feedback and related complaints about product features, and the corresponding tags are "product feature feedback" and "customer complaint risk," then the generated summary might be "The customer reported that a certain product feature is inconvenient to use and expressed dissatisfaction with the product, indicating a risk of customer complaint."

[0093] Step 302: Obtain the tag statistics of the multiple target tags, including the occurrence frequency of each target tag and the time period distribution of each target tag.

[0094] Step 303: In response to the user's selection of a preset output chart type, output a visual statistical chart of the label statistics information.

[0095] In the two steps above, tag statistics are collected and organized. This information includes the frequency of different tags and their distribution over different time periods. Then, for example, the Matplotlib plotting library can be used. Users can choose the appropriate chart type based on their specific needs. For instance, to show the proportion of different tags in all calls, a pie chart can be drawn, with the size of each sector representing the percentage of the corresponding tag. To observe the trend of a tag's quantity over a period of time, a bar chart can be drawn, with time on the horizontal axis and the number of tags appearing on the vertical axis. During the plotting process, the charts are also enhanced, including adding titles, axis labels, and legends to make them more intuitive, clear, and easier for users to understand.

[0096] Step 304: Based on the call summary and the visual statistical charts, obtain an analysis report of the call content, wherein the analysis report conforms to a preset report template.

[0097] For example, the preset report template typically uses HTML, Markdown, or other formats, defining the overall structure and layout of the report, including the title, summary section, and chart display area. The call content summary generated in step 301 is filled into the corresponding positions in the template, and the visualization charts generated in step 303 are embedded into the template in an appropriate manner. In this way, textual and visual information are organically combined to form a structured data report, making the report content rich, organized, and easy for users to view and analyze.

[0098] For example, in this embodiment of the application, the generated structured data report can be converted into different formats for export using appropriate tools and libraries according to user needs. For example, to export to HTML format, the generated HTML code can be saved directly; to export to PDF format, tools such as pdfkit can be used to convert the HTML report into a PDF file; for Word format, the python-docx library can be used to reorganize the report content according to the structure requirements of a Word document and save it as a .docx file. In this way, users can flexibly choose the appropriate report format according to their actual usage scenarios and needs.

[0099] For example, the obtained call content analysis report is stored in the enterprise's internal database and can be indexed for easy export. At the same time, the example provided in this application also provides a download interface to support users to download the report as needed, and records the download log.

[0100] For example, choose a suitable internal enterprise database system, such as the relational database MySQL or the non-relational database MongoDB. Establish a connection to the database according to the database connection method and operating procedures. Insert the generated data report into the database according to a specific format, such as JSON format for MongoDB, or according to the designed table structure for MySQL. Simultaneously, to improve the efficiency of subsequent queries, create indexes for key information in the report, such as the report generation time and call identifiers. This way, when a specific report needs to be queried, the relevant record can be quickly located through the index, reducing query time.

[0101] For example, a corresponding download interface is developed, such as a web-based API, allowing users to download reports via network requests. The download interface receives user request parameters, such as the report identifier, and then retrieves the corresponding report data from the database based on these parameters. The report data is returned to the user in a suitable format, such as the requested HTML, PDF, or Word format, for download. While the user is downloading the report, a download log is recorded, including download time, the downloading user's identifier, and the downloaded report identifier. This log information can be stored in a dedicated log table in the database for subsequent statistical analysis of report downloads and to understand user usage of the reports.

[0102] For example, unlike general analysis models in fields such as power and e-commerce, the embodiments of this application focus on the real-time nature of call content, speech-to-text characteristics, and business relevance, and can be quickly adapted to scenarios such as customer service, financial risk control, and security monitoring without retraining the basic model.

[0103] The following details how to obtain the label classification results of the multiple target labels based on the multiple target labels and multiple pre-set similarity thresholds. Two exemplary methods are provided in this application embodiment.

[0104] The first exemplary method:

[0105] The method for obtaining label classification results for the multiple target labels based on the multiple target labels and multiple pre-set similarity thresholds includes:

[0106] Based on the comparison between the semantic similarity of the target tags and the preset multiple similarity thresholds, the multiple target tags are classified into primary tags, secondary tags and related tags. The second similarity threshold corresponding to the primary tag is greater than the third similarity threshold corresponding to the secondary tag, and the second similarity threshold corresponding to the secondary tag is greater than the first similarity threshold corresponding to the related tag.

[0107] For example, in this exemplary method, matching target tags are directly searched for in a pre-defined tag library based on semantic similarity for classification. The primary tags are those that are highly similar semantically and have high importance. The corresponding second similarity threshold is the highest, meaning that only when the semantic similarity between two tags exceeds this high threshold will they be classified as primary tags. Primary tags typically directly reflect the core theme or key features of the data and are the most crucial part of data classification and retrieval.

[0108] For example, secondary tags have lower semantic similarity than primary tags but higher semantic similarity than related tags. Their corresponding third similarity threshold is lower than the second similarity threshold of the primary tag but still higher than the first similarity threshold of the related tag. Although secondary tags are slightly less important than primary tags, they can still provide supplementary information related to the primary tag, helping to more comprehensively describe the details and context of the data.

[0109] For example, related tags are those tags with the lowest semantic similarity. Their semantic similarity is just above the lowest first similarity threshold, indicating that they are related to the primary or secondary tags to some extent, but the similarity is low. Related tags can expand the semantic scope of data and help users discover other potential information related to the core topic, but they are not the core features of the data themselves.

[0110] For example, when a call text corresponds to multiple tags, multi-tag fusion processing can be performed, using statistical theories of attribution analysis and cluster analysis.

[0111] This classification method, based on semantic similarity and similarity thresholds, enables the system to manage target tags hierarchically. Primary tags can be used directly to quickly locate and identify the core content of the data; secondary tags provide richer details, helping users gain a deeper understanding of the data's background and characteristics; and related tags can broaden the user's perspective, uncovering other potential information related to the core topic. This method not only improves data organization efficiency but also facilitates subsequent call content analysis, thereby enhancing the accuracy of call content analysis.

[0112] The second exemplary method:

[0113] In this second exemplary method, in order to identify potential tag associations, the tags in the preset tag library are first clustered according to their correlation with each other. Based on this, the tag classification results of the multiple target tags in the selected association clusters are further obtained according to semantic similarity.

[0114] The semantic similarity is used to obtain multiple target labels corresponding to the semantic feature vector, such as... Figure 4 As shown, the method includes the following steps:

[0115] Step 401: Based on the tag similarity between the first tag and the second tag in the preset tag library, cluster the tags in the preset tag library, where the first tag and the second tag are both any tags in the preset tag library.

[0116] For example, in this step, tags in a preset tag library are clustered. Clustering is based on the similarity between tags. Specifically, the clustering of tags within the preset tag library is based on the tag similarity between any two tags in the preset tag library, such as a first tag and a second tag. For example, tag similarity can be based on word co-occurrence frequency, semantic relevance, or other relevance indicators.

[0117] Through clustering, all tags in the tag library are divided into different clusters, with tags within each cluster having a high semantic relevance. This process is fundamental to tag classification, as it initially groups a large number of tags according to semantic similarity, providing a structured framework for subsequent classification.

[0118] Step 402: Based on the clustering results and the preset tag similarity threshold, all tags in the preset tag library are divided into high-association clusters, medium-association clusters, and low-association clusters. Each tag contained in the same association cluster is semantically related.

[0119] For example, clustering uses the similarity between labels as a metric to group semantically similar labels into the same cluster. For instance, "Analysis of reasons for order not being picked up on time" and "delivery timeliness issues" may be clustered into the "order fulfillment" cluster.

[0120] Based on the clustering results and a preset label similarity threshold, all labels are further divided into high-association clusters, medium-association clusters, and low-association clusters. The label similarity threshold is used to distinguish clusters with different degrees of association. Labels in high-association clusters have the highest semantic similarity and usually represent the most core and closely related parts of the label library; labels in medium-association clusters have the next highest similarity and reflect relatively obvious semantic associations; while low-association clusters contain labels that are relatively independent in meaning but still have some connection.

[0121] For example, all tags in the preset tag library can be divided into any number of clusters, depending on the number of clusters determined in the clustering process according to requirements.

[0122] For example, the clustering method can be the KMeans clustering algorithm, in which the similarity values ​​of all labels are converted into a two-dimensional array and used as input to KMeans. For example, if the number of clusters n_clusters is set to 3, then the labels are divided into high, medium and low association clusters according to the small similarity of the labels.

[0123] Step 403: Obtain the target association clusters corresponding to the labels whose semantic similarity is greater than the first similarity threshold as multiple target labels, wherein the target association cluster is at least one of the high association cluster, medium association cluster and low association cluster.

[0124] For example, each association cluster contains multiple semantically related labels, so the target association cluster can be determined by comparing the average semantic similarity of the association cluster with a first similarity threshold.

[0125] For example, the highly correlated clusters are: [“Analysis of reasons for order not being picked up in time”, “Delivery timeliness issues”], where the average semantic similarity of the highly correlated clusters is 0.82.

[0126] Medium-interrelated clusters: ["Employee service evaluation", "Customer complaint risk"]. The average semantic similarity of these medium-interrelated clusters is 0.68. The first similarity threshold can be either 0.68 or 0.82. If it is 0.68, then the target related cluster falls into the high-interrelated cluster and the medium-interrelated cluster.

[0127] For example, the target association cluster can also be determined by the proportion of labels with semantic similarity greater than the first similarity threshold in the association cluster. For example, if the proportion of multiple labels in a high association cluster whose similarity to the semantic feature vector meets the requirements is 80%, exceeding the preset proportion threshold of 70%, then the high association cluster can be identified as the target association cluster, and all the labels in the target association cluster are multiple target labels.

[0128] For example, the target association cluster is a cluster whose semantic similarity is greater than a first similarity threshold, and can be one or more of high association clusters, medium association clusters, or low association clusters.

[0129] For example, the cluster with the highest similarity to the semantic vector of the current call can also be extracted, the highly correlated cluster, in which the labels are highly correlated with the call content.

[0130] After obtaining the pre-defined label library after clustering, the following details how to obtain the label classification results of the multiple target labels within the pre-defined label library after clustering.

[0131] For example, the step of obtaining the label classification results of the multiple target labels based on the multiple target labels and multiple pre-set similarity thresholds is as follows: Figure 5 As shown, the method includes the following steps:

[0132] Step 501: If the target association cluster includes both high association clusters and medium association clusters, select the tags with semantic similarity greater than the second similarity threshold from the high association clusters as the main tags.

[0133] For example, in this step, the labels in the target association cluster are classified. If the target association cluster contains high-association clusters and medium-association clusters, labels with semantic similarity greater than a second similarity threshold are selected from the high-association clusters as primary labels. Primary labels are the most important part of label classification; they represent the core features or themes of the data. Since the labels in the high-association clusters already have high semantic similarity, further filtering through the second similarity threshold ensures the high quality and high relevance of the primary labels.

[0134] Step 502: Select tags from the high-association clusters and medium-association clusters whose semantic similarity is less than or equal to the second similarity threshold and greater than the first similarity threshold as secondary tags.

[0135] In this step, labels with semantic similarity less than or equal to the second similarity threshold but greater than the first similarity threshold are selected from the high-relevance clusters and medium-relevance clusters as secondary labels. Although secondary labels are slightly less important than primary labels, they still provide supplementary information related to the primary labels, helping to more comprehensively describe the details and context of the data. Through this intermediate-level similarity threshold screening, secondary labels can balance the high relevance of primary labels and the low relevance of labels in low-relevance clusters, providing a richer semantic hierarchy for data classification.

[0136] Step 503: Select other tags from the highly correlated clusters as correlated tags. The other tags are tags that do not include the primary tag and the secondary tag.

[0137] In this step, other tags are selected as associated tags from the high-relevance cluster. These associated tags are those not included in the primary and secondary tags. Although they are semantically related to the primary and secondary tags, their similarity is low, and therefore they are classified as associated tags. Because these associated tags are obtained from the high-relevance cluster, they are also closely related to the content of the call.

[0138] The following section details how to obtain the semantic feature vector corresponding to the call text.

[0139] In this embodiment, semantic feature vectors are obtained by combining a bidirectional long short-term memory network (BiLSTM) neural network architecture with an attention mechanism. The detailed process of this method is as follows.

[0140] For example, such as Figure 6 As shown, the neural network model is obtained in advance through the following methods:

[0141] Step 601: Construct a neural network architecture based on a bidirectional long short-term memory network (BiLSTM). The neural network architecture includes an input layer, an embedding layer, a BiLSTM layer, and an output layer.

[0142] For example, a BiLSTM network structure can be built using deep learning frameworks such as TensorFlow or PyTorch, defining the input layer, embedding layer, BiLSTM layer, output layer, etc.

[0143] For example, the input layer is used to receive a numeric vector mapped by the vocabulary; the embedding layer is used to convert the numeric vector into a word embedding vector; the BiLSTM layer is used to process the sequence output by the embedding layer, extract the forward and backward features at each time step, and output the feature vector at each time step; the attention mechanism layer is used to perform weighted aggregation on the feature vectors of all time steps output by the BiLSTM layer to obtain a fixed-length context vector, where the weights represent the importance of each time step; the output layer is used to generate the semantic feature vector based on the context vector output by the attention mechanism layer.

[0144] Step 602: Add an attention mechanism layer to the neural network architecture. The attention mechanism layer is placed after the BiLSTM layer and before the output layer. It is used to perform weighted aggregation of the temporal features extracted by the BiLSTM layer.

[0145] In step 602, the attention mechanism layer is configured between the BiLSTM layer and the output layer to perform weighted aggregation processing on the temporal features output by the BiLSTM layer. The attention mechanism layer evaluates the importance of each time step feature extracted by the BiLSTM layer through an adaptive weight allocation mechanism, assigning higher weights to features representing key semantic information and lower weights to secondary features, thereby generating a weighted aggregation of core semantic feature representations. This core semantic feature representation is passed to the output layer to perform a label classification task, enabling the model to prioritize core feature combinations related to the classification target when processing call text containing scattered key information, thereby improving the accuracy of enumerated value classification.

[0146] Existing technologies focus on neural network hardware adaptation and do not involve semantic analysis of call content. Traditional LSTM models only process sequences in one direction and have weak semantic understanding capabilities for complex sentences. However, the embodiments of this application can obtain more accurate semantic feature vector extraction through BiLSTM + attention mechanism.

[0147] For example, obtaining the semantic feature vector corresponding to the call text, such as Figure 7 As shown, the method includes the following steps:

[0148] Step 701: Map the call text into a number vector according to a preset vocabulary.

[0149] For example, this is usually done by analyzing all the training text data, segmenting and removing stop words, counting all the words that have appeared, and then selecting the N most frequent words or all words and arranging them in a certain order, such as descending frequency or alphabetical order, to form a dictionary.

[0150] In this step, each word in the text is converted into a unique identifier corresponding to the vocabulary, thereby improving the resource utilization of the model.

[0151] Step 702: Input the numerical vector corresponding to the call text into a pre-obtained neural network model to obtain the semantic feature vector corresponding to the call text, wherein the neural network model is trained by training the call text and the training labels corresponding to the training call text.

[0152] For example, the neural network model is a BiLSTM-based neural network model combined with an attention mechanism, built using the TensorFlow framework. For example, during training, 100,000 past call texts and corresponding tag data from the enterprise are collected and divided into a training set, a validation set, and a test set in a 7:1:2 ratio, where:

[0153] The training set is used to adjust the weights of the neural network and minimize the loss function through the backpropagation algorithm.

[0154] The validation set is used for hyperparameter tuning, such as adjusting the learning rate, batch size, and network structure, to prevent overfitting and for model selection.

[0155] The test set is used to evaluate the model's generalization ability; its ability to handle new data reflects the model's expected performance in real-world scenarios. For example, the learning rate can be set to 0.001, the number of hidden layer neurons to 128, and the number of training epochs to 30. The training set is input into the model for training, and the parameters are adjusted based on the validation set results. After training, the preprocessed real-time call text is input into the model to extract semantic feature vectors.

[0156] For example, during the training process, call text data labeled with tags, such as sentiment tags and business category tags, is divided into training sets and validation sets, which are then input into the built neural network model. The model parameters are updated through the backpropagation algorithm, enabling the model to learn the mapping relationship between text and tags.

[0157] For example, during neural network training, parameter updates occur in the iterations of forward and backward propagation on the training set. Essentially, this is a process of self-correction through repeated trial and error. In detail, when text data from the training set is input into the neural network model, forward propagation is performed: the input data is passed layer by layer from the embedding layer, passes through a bidirectional LSTM to extract contextual features, then focuses on keywords through an attention mechanism, and finally outputs the prediction result. At this point, the loss function, such as the cross-entropy loss function, calculates the degree of difference between the predicted label and the true label; this difference value is the model's current "error signal."

[0158] Next, during the backpropagation phase, this error signal flows backward along the network—from the output layer back to the input layer layer by layer. In this process, the system uses the chain rule to efficiently calculate the partial derivative of the loss function with respect to each parameter, i.e., the gradient. For example, in an LSTM layer, the gradient is precisely decomposed into the weight matrices of the input gate, forget gate, and output gate, as well as parameters related to the cell state, revealing the specific contribution of each parameter to the total error. Finally, after obtaining the gradients of all parameters, the optimizer (such as Adam) immediately performs parameter updates: subtracting the product of the learning rate (0.001) and the gradient from the current parameter value. This is equivalent to moving the model a small step along the steepest descent direction of the error surface. The entire process iterates on each batch of training data: input text—error calculation—backpropagation—gradient update. After multiple corrections, the model parameters gradually converge to the optimal state, ultimately establishing a precise mapping relationship from call text to business tags.

[0159] Finally, the numerical vector corresponding to the call text is input into a pre-obtained neural network model, and the semantic feature vector corresponding to the call text is obtained through this mapping relationship for subsequent similarity calculation.

[0160] It should be understood that, although Figure 2-7 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order in which these steps are executed, and they can be performed in other orders. Figure 2-7 At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0161] In some embodiments, such as Figure 8As shown, a call content analysis device is provided, the device comprising: an acquisition module 801, a processing module 802, and an analysis module 803, wherein:

[0162] The acquisition module 801 is used to obtain the semantic feature vectors corresponding to the call text between the customer service representative and the customer in the call scenario.

[0163] The processing module 802 is used to obtain the semantic similarity between the semantic feature vector and each tag vector in the preset tag library; obtain multiple target tags corresponding to the semantic feature vector based on the semantic similarity; and obtain the tag classification results of the multiple target tags based on the multiple target tags and multiple preset similarity thresholds, wherein the tag classification results include primary tags, secondary tags and related tags.

[0164] Analysis module 803 is used to obtain an analysis report of the call text based on the tag classification results.

[0165] Further limitations regarding the call content analysis device can be found in the limitations of the call content analysis method described above, and will not be repeated here. Each module in the aforementioned call content analysis device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in the terminal device in hardware form, or stored in the memory of the terminal device in software form, so that the processor can call and execute the corresponding operations of each module.

[0166] Another embodiment provides a computer-readable storage medium for storing a computer program. This computer program contains instructions for implementing the methods described in the embodiments of this application. By installing this computer program on a computer, the computer can execute the corresponding methods.

[0167] Another embodiment proposes a computer program product that includes computer program code. When this computer program code is run on a computer, it causes the computer to implement the methods proposed in the embodiments of this application. Thus, a user can implement these methods by using this computer program product.

[0168] For example, Figure 9 This is a schematic block diagram of the electronic device provided in the embodiments of this application.

[0169] Electronic device 900 may include: a memory 901 storing executable program code and a processor 902 coupled to the memory 901.

[0170] In this embodiment, processor 902 calls executable program code stored in memory to execute any of the methods disclosed in the embodiments of this application. Those skilled in the art will understand that... Figure 9 The electronic device structure shown does not constitute a limitation on the electronic device. The electronic device may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0171] The processor 902 is the control center of the electronic device. It connects various parts of the electronic device via various interfaces and lines, and performs various functions and processes data by running or executing software programs and / or modules stored in the memory, and by calling data stored in the memory, thereby providing overall monitoring of the electronic device. Optionally, the processor may include one or more processing units; preferably, the processor may integrate an application processor and a modem processor, wherein the application processor mainly handles the operating system, user interface, and applications, and the modem processor mainly handles wireless communication. It is understood that the modem processor may also not be integrated into the processor.

[0172] The memory 901 can be used to store software programs and modules. The processor executes various functional applications and data processing of the electronic device by running the software programs and modules stored in the memory. The memory may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function, etc.; the data storage area may store data created according to the use of the electronic device, etc. In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0173] It should be understood that, in the embodiments of this application, the processor may be a Central Processing Unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor, etc.

[0174] In implementation, each step of the above method can be completed by integrated logic circuits in the processor's hardware or by instructions in software. The steps of the method disclosed in the embodiments of this application can be directly manifested as execution by a hardware processor, or as a combination of hardware and software modules within the processor. The software modules can reside in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. This storage medium is located in memory, and the processor executes the instructions in the memory, combining them with its hardware to complete the steps of the above method. To avoid repetition, detailed descriptions are omitted here.

[0175] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the embodiments of this application.

[0176] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0177] In the embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative. For instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual couplings, direct couplings, or communication connections may be through some interfaces; indirect couplings or communication connections between devices or units may be electrical, mechanical, or other forms.

[0178] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0179] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0180] If the aforementioned function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application embodiment, essentially, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0181] The above description is merely a specific implementation of the embodiments of this application, but the protection scope of the embodiments of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the embodiments of this application should be included within the protection scope of the embodiments of this application. Therefore, the protection scope of the embodiments of this application should be determined by the protection scope of the claims.

Claims

1. A call content analysis method characterized by, The method includes: Obtain the semantic feature vectors corresponding to the text of the call between the customer service representative and the customer in the call scenario; Obtain the semantic similarity between the semantic feature vector and each tag vector in the preset tag library; Based on the semantic similarity, multiple target labels corresponding to the semantic feature vector are obtained; Based on the multiple target tags and multiple pre-set similarity thresholds, the tag classification results of the multiple target tags are obtained, and the tag classification results include primary tags, secondary tags and related tags; Based on the tag classification results, an analysis report of the call text is obtained; The step of obtaining multiple target labels corresponding to the semantic feature vector based on the semantic similarity includes: Based on the similarity between the first tag and the second tag in the preset tag library, the tags in the preset tag library are clustered, where the first tag and the second tag are both any tags in the preset tag library; Based on the clustering results and the preset tag similarity threshold, all tags in the preset tag library are divided into high-association clusters, medium-association clusters, and low-association clusters, and each tag contained in the same association cluster is semantically related. The target association clusters corresponding to the labels whose semantic similarity is greater than the first similarity threshold are obtained as multiple target labels, and the target association clusters are at least one of the high association cluster, medium association cluster and low association cluster; The step of obtaining the label classification results of the multiple target labels based on the multiple target labels and multiple pre-set similarity thresholds includes: When the target association cluster includes both high association clusters and medium association clusters, the labels with semantic similarity greater than the second similarity threshold are selected from the high association clusters as the main labels; From the highly correlated clusters and the moderately correlated clusters, select the tags whose semantic similarity is less than or equal to the second similarity threshold and greater than the first similarity threshold as secondary tags; Other tags are selected from the highly correlated clusters as correlated tags, wherein the other tags are tags that do not include the primary tag and the secondary tag.

2. The method of claim 1, wherein, The step of obtaining the label classification results of the multiple target labels based on the multiple target labels and multiple pre-set similarity thresholds includes: Based on the comparison between the semantic similarity of the target tags and the preset multiple similarity thresholds, the multiple target tags are classified into primary tags, secondary tags and related tags. The second similarity threshold corresponding to the primary tag is greater than the third similarity threshold corresponding to the secondary tag, and the second similarity threshold corresponding to the secondary tag is greater than the first similarity threshold corresponding to the related tag.

3. The method of claim 1, wherein, Obtaining the semantic feature vector corresponding to the call text includes: The call text is mapped into a number vector according to a preset vocabulary; The numerical vector corresponding to the call text is input into a pre-obtained neural network model to obtain the semantic feature vector corresponding to the call text, wherein the neural network model is obtained by training the call text and the training labels corresponding to the training call text.

4. The method of claim 3, wherein, The neural network model is obtained in advance through methods including: A neural network architecture based on Bidirectional Long Short-Term Memory (BiLSTM) is constructed, comprising an input layer, an embedding layer, a BiLSTM layer, and an output layer. An attention mechanism layer is added to the neural network architecture. This attention mechanism layer is placed after the BiLSTM layer and before the output layer. It is used to perform weighted aggregation of the temporal features extracted by the BiLSTM layer.

5. The method according to claim 1, characterized in that, The step of obtaining an analysis report of the call text based on the tag classification results includes: Based on the call text, multiple target tags corresponding to the call text, and the preset weights of the primary tags, secondary tags, and associated tags among the multiple target tags, a call summary corresponding to the call content is obtained; Obtain tag statistics information for the multiple target tags, including the frequency of occurrence of each target tag and the time period distribution of each target tag; In response to the user's selection of a preset output chart type, a visual statistical chart of the label statistics information is output. An analysis report of the call content is obtained based on the call summary and the visual statistical charts, and the analysis report conforms to a preset report template.

6. A call content analysis device, characterized in that, The device includes: The acquisition module is used to obtain the semantic feature vectors corresponding to the call text between customer service and customer in a call scenario; The processing module is used to obtain the semantic similarity between the semantic feature vector and each tag vector in the preset tag library; obtain multiple target tags corresponding to the semantic feature vector based on the semantic similarity; and obtain the tag classification results of the multiple target tags based on the multiple target tags and multiple preset similarity thresholds, wherein the tag classification results include primary tags, secondary tags and related tags. The analysis module is used to obtain an analysis report of the call text based on the tag classification results; The step of obtaining multiple target labels corresponding to the semantic feature vector based on the semantic similarity includes: clustering the labels in the preset label library based on the label similarity between the first label and the second label in the preset label library, wherein the first label and the second label are both any labels in the preset label library; dividing all labels in the preset label library into high-association clusters, medium-association clusters, and low-association clusters based on the clustering results and a preset label similarity threshold, wherein each label contained in the same association cluster is semantically related; and obtaining the target association clusters corresponding to the labels whose semantic similarity is greater than the first similarity threshold as multiple target labels, wherein the target association cluster is at least one of the high-association cluster, medium-association cluster, and low-association cluster. The step of obtaining the tag classification results of the multiple target tags based on the multiple target tags and multiple pre-set similarity thresholds includes: when the target association cluster includes a high association cluster and a medium association cluster, selecting tags from the high association cluster whose semantic similarity is greater than a second similarity threshold as primary tags; selecting tags from the high association cluster and the medium association cluster whose semantic similarity is less than or equal to the second similarity threshold and greater than the first similarity threshold as secondary tags; and selecting other tags from the high association cluster as associated tags, wherein the other tags are tags that do not include primary tags and secondary tags.

7. An electronic device, characterized in that, It includes a processor and a memory, the memory storing a computer program, and the processor executing the computer program to implement the steps of the method according to any one of claims 1-5.

8. A computer-readable storage medium, characterized in that, The device stores computer instructions that, when executed by a processor, implement the steps of the method according to any one of claims 1-5.

Images