Event mining method and system, electronic equipment and storage medium

Abstract

The embodiment of the invention provides an event mining method and system, electronic equipment and a storage medium, and belongs to the technical field of artificial intelligence. The method comprises the steps of obtaining a target customer complaint text, wherein the target customer complaint text is from a customer complaint dialogue; performing event extraction on the target customer complaint text based on a pre-trained event generation model to obtain candidate events; screening out a reference event from a preset event library based on the candidate event; performing similarity detection on the candidate event and the reference event based on a pre-trained first event detection model to obtain an event similarity result; and if the event similarity result represents that the candidate event is not similar to the reference event, performing event screening on the candidate event based on a pre-trained second event detection model and the target customer complaint text to obtain a target mining event. According to the embodiment of the invention, the event mining accuracy can be improved.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular, to an event mining method and system, an electronic device, and a storage medium. Background Art

[0002] Event mining is a relatively complex problem in different application scenarios, especially in the customer complaint dialogue scenario, mining the problems raised by customers and the solutions to the problems. For example, if the problems raised by customers are unauthorized signing for express delivery and package loss, the corresponding solutions are to claim compensation or file a complaint. Currently, related technologies often construct event mining by combining deep learning mining with manual verification. However, in the case of a large amount of data or unstructured data scenarios, this method will rely on the understanding of the scenario by the person who constructs it, and at the same time requires a large amount of manpower and resources. Moreover, manual annotation is prone to missing or mislabeling. Therefore, due to the high degree of dependence on people, the accuracy of event mining is not high. Summary of the Invention

[0003] The main purpose of the embodiments of this application is to propose an event mining method and system, an electronic device, and a storage medium, aiming to improve the accuracy of event mining.

[0004] To achieve the above object, in the first aspect of the embodiments of this application, an event mining method is proposed. The method includes: obtaining a target customer complaint text, where the target customer complaint text is derived from a customer complaint dialogue;

[0005] Performing event extraction on the target customer complaint text based on a pre-trained event generation model to obtain candidate events;

[0006] Screening out reference events from a preset event library based on the candidate events;

[0007] Performing similarity detection on the candidate events and the reference events based on a pre-trained first event detection model to obtain an event similarity result;

[0008] If the event similarity result indicates that the candidate event is not similar to the reference event, then performing event screening on the candidate event based on a pre-trained second event detection model and the target customer complaint text to obtain a target mining event.

[0009] In some embodiments, after performing similarity detection on the candidate events and the reference events based on a pre-trained first event detection model to obtain an event similarity result, the method further includes:

[0010] If the event similarity result indicates that the candidate event is similar to the reference event, then performing a merging process on the candidate event and the reference event to obtain a merged event;

[0011] Replace the reference event with the merged event.

[0012] In some embodiments, the merging process of the candidate event and the reference event to obtain a merged event includes:

[0013] Extract keywords from the candidate event to obtain candidate keywords;

[0014] Extract keywords from the reference event to obtain reference keywords;

[0015] Calculate the distance between the candidate keywords and the reference keywords to obtain the keyword feature distance between the candidate keywords and the reference keywords;

[0016] Construct an event graph based on the candidate keywords, the reference keywords, and the keyword feature distance; wherein, the candidate keywords and the reference keywords serve as nodes of the event graph, and the keyword feature distance serves as an edge of the event graph;

[0017] Perform clustering calculation on the nodes in the event graph to obtain at least one keyword node group;

[0018] Reconstruct sentences for each keyword node group based on the reference event and the candidate event to obtain the merged event.

[0019] In some embodiments, the screening of the reference event from the preset event library based on the candidate event includes:

[0020] Determine target keywords based on the candidate event;

[0021] Screen events from the preset event library based on the target keywords to obtain the reference event.

[0022] In some embodiments, the pre-trained second event detection model includes a preprocessing layer, a Transformer layer, a pooling layer, and a fully connected layer. The screening of the candidate event based on the pre-trained second event detection model and the target customer complaint text to obtain a target mining event includes:

[0023] Perform word segmentation and embedding processing on the candidate event based on the preprocessing layer to obtain a word vector sequence;

[0024] Perform encoding processing on the word vector sequence based on the Transformer layer to obtain a hidden state sequence;

[0025] Extract information from the hidden state sequence based on the pooling layer to obtain a first semantic information vector;

[0026] Obtain the semantic information vector of the target customer complaint text to obtain a second semantic information vector;

[0027] Based on the fully connected layer and the second semantic information vector, classify the first semantic information vector to obtain a probability distribution result;

[0028] Based on the probability distribution result, perform event screening on the candidate events to obtain the target mined event.

[0029] In some embodiments, the event generation model is constructed in the following manner:

[0030] Generate a prompt according to the event mining scenario to obtain a prompt text;

[0031] Based on the prompt text and a predetermined large language model, perform model construction to obtain an event generation model.

[0032] In some embodiments, the target customer complaint text is obtained in the following manner:

[0033] Obtain the audio data of the customer complaint conversation;

[0034] Perform conversion processing on the audio data of the customer complaint conversation to obtain the original customer complaint text;

[0035] Perform correction processing on the original customer complaint text to obtain an intermediate customer complaint text;

[0036] Perform replacement processing on the intermediate customer complaint text through a preset term library to obtain the target customer complaint text.

[0037] To achieve the above object, a second aspect of the embodiments of the present application proposes an event mining system, and the system includes:

[0038] A data acquisition module, configured to acquire a target customer complaint text, and the target customer complaint text is derived from a customer complaint conversation;

[0039] A candidate event acquisition module, configured to perform event extraction on the target customer complaint text based on a pre-trained event generation model to obtain candidate events;

[0040] A reference event acquisition module, configured to screen out reference events from a preset event library based on the candidate events;

[0041] A first detection module, configured to perform similarity detection on the candidate events and the reference events based on a pre-trained first event detection model to obtain an event similarity result;

[0042] A second detection module, configured to, if the event similarity result indicates that the candidate event is not similar to the reference event, perform event screening on the candidate event based on a pre-trained second event detection model and the target customer complaint text to obtain a target mining event.

[0043] To achieve the above object, a third aspect of the embodiments of the present application provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, the method described in the first aspect above is implemented.

[0044] To achieve the above object, a fourth aspect of the embodiments of the present application provides a computer-readable storage medium, which stores a computer program, and when the computer program is executed by a processor, the method described in the first aspect above is implemented.

[0045] The event mining method, system, electronic device and storage medium provided by the present application obtain a target customer complaint text, which is derived from a customer complaint conversation; perform event extraction on the target customer complaint text based on a pre-trained event generation model to obtain candidate events, and can initially obtain events from the target customer complaint text. Compared with manual extraction, the event generation model can be more efficient. At the same time, extracting candidate events from the target customer complaint text provides a structured input for subsequent steps, which is conducive to improving the matching efficiency and accuracy. Further, based on the candidate events, reference events are screened out from a preset event library. Compared with traversing the event library to match the events in the event library with the candidate events one by one, the pre-screening operation can filter out irrelevant events in the event library, reduce the matching calculation amount, and improve the accuracy of event mining. Further, based on a pre-trained first event detection model, similarity detection is performed on the candidate event and the reference event to obtain an event similarity result, which can identify whether the candidate event already exists in the event library, thereby avoiding redundant information in the event library. Further, if the event similarity result indicates that the candidate event is not similar to the reference event, event screening is performed on the candidate event based on a pre-trained second event detection model and the target customer complaint text to obtain a target mining event. By performing semantic screening on the dissimilar candidate events based on the target customer complaint text, the accuracy of candidate event generation is improved, thereby improving the accuracy of event mining. Description of the Drawings

[0046] Figure 1 is a flowchart of the event mining method provided by the embodiments of the present application;

[0047] Figure 2 is Figure 1 a flowchart of step S103 in

[0048] Figure 3 It is the principle flowchart of the first event detection model in the event mining method provided by the embodiments of the present application;

[0049] Figure 4 It is another flowchart of the event mining method provided by the embodiments of the present application;

[0050] Figure 5 It is another flowchart of the event mining method provided by the embodiments of the present application;

[0051] Figure 6 It is another flowchart of the event mining method provided by the embodiments of the present application;

[0052] Figure 7 It is a schematic diagram of the specific implementation process of the event mining method provided by the embodiments of the present application;

[0053] Figure 8 It is a schematic diagram of the structure of the event mining system provided by the embodiments of the present application;

[0054] Figure 9 It is a schematic diagram of the hardware structure of the electronic device provided by the embodiments of the present application. Detailed implementation manners

[0055] In order to make the objectives, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application, and are not used to limit the present application.

[0056] It should be noted that although functional module division is performed in the system schematic diagram and the logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order from the module division in the system or the order in the flowchart. Terms such as "first" and "second" in the specification, claims and the above-mentioned drawings are used to distinguish similar objects, and do not have to be used to describe a specific order or sequence.

[0057] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the technical field to which the present application belongs. The terms used herein are only for the purpose of describing the embodiments of the present application, and are not intended to limit the present application.

[0058] First, several nouns involved in the present application are analyzed:

[0059] Artificial Intelligence (AI): It is a new technical science that studies and develops theories, methods, technologies, and application systems for simulating, extending, and expanding human intelligence; artificial intelligence is a branch of computer science. Artificial intelligence attempts to understand the essence of intelligence and produce a new intelligent machine that can react in a way similar to human intelligence. The research in this field includes robots, speech recognition, image recognition, natural language processing, and expert systems, etc. Artificial intelligence can simulate the information process of human consciousness and thinking. Artificial intelligence also refers to the theory, method, technology, and application system that uses a digital computer or a machine controlled by a digital computer to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use knowledge to obtain the best results.

[0060] Natural Language Processing (NLP): NLP uses computers to process, understand, and apply human languages (such as Chinese, English, etc.). NLP is a branch of artificial intelligence and an interdisciplinary field of computer science and linguistics, and is often referred to as computational linguistics. Natural language processing includes syntactic analysis, semantic analysis, discourse understanding, etc. Natural language processing is commonly used in technical fields such as machine translation, handwritten and printed character recognition, speech recognition and text-to-speech conversion, information intention recognition, information extraction and filtering, text classification and clustering, public opinion analysis, and opinion mining. It involves data mining, machine learning, knowledge acquisition, knowledge engineering, artificial intelligence research related to language processing, and linguistic research related to language computing.

[0061] Information Extraction: A text processing technology that extracts factual information such as entities, relationships, events of a specified type from natural language texts and forms structured data for output. Information extraction is a technology for extracting specific information from text data. Text data is composed of some specific units, such as sentences, paragraphs, and passages. Text information is exactly composed of some small specific units, such as characters, words, phrases, sentences, paragraphs, or combinations of these specific units. Extracting noun phrases, personal names, place names, etc. from text data are all text information extraction. Of course, the information extracted by text information extraction technology can be various types of information.

[0062] Event mining is a relatively complex problem in different application scenarios, especially in the scenario of customer complaint conversations, where it is necessary to mine the problems raised by customers and the solutions to the problems. For example, if the problems raised by customers are unauthorized signing for express delivery and package loss, the corresponding solutions are to claim compensation or lodge a complaint. Currently, relevant technologies often construct event mining by combining deep learning mining with manual verification. However, in the case of a large amount of data or unstructured data scenarios, this method will rely on the understanding of the scenario by the person constructing it, and at the same time requires a large amount of manpower and resource investment, and manual annotation is prone to missing or mislabeling. Therefore, due to the high degree of dependence on people, the accuracy of event mining is not high.

[0063] Based on this, the embodiments of the present application provide an event mining method, system, electronic device and storage medium, aiming to improve the accuracy of event mining.

[0064] The event mining method, system, electronic device and storage medium provided by the embodiments of the present application are specifically described through the following embodiments. First, the event mining method in the embodiments of the present application is described.

[0065] The embodiments of the present application can acquire and process relevant data based on artificial intelligence technology. Among them, artificial intelligence (AI) is a theory, method, technology and application system that uses digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results.

[0066] Artificial intelligence basic technologies generally include technologies such as sensors, dedicated artificial intelligence chips, cloud computing, distributed storage, big data processing technology, operation / interaction systems, and mechatronics. Artificial intelligence software technologies mainly include several major directions such as computer vision technology, robotics, biometric technology, speech processing technology, natural language processing technology, and machine learning / deep learning.

[0067] The event mining method provided by the embodiments of the present application relates to the field of artificial intelligence technology. The event mining method provided by the embodiments of the present application can be applied to terminals, or to server sides, or can also be software running on terminals or server sides. In some embodiments, the terminal can be a smart phone, a tablet computer, a notebook computer, a desktop computer, etc.; the server side can be configured as an independent physical server, or can be configured as a server cluster or distributed system composed of multiple physical servers, or can also be configured as a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms; the software can be an application that implements the event mining method, etc., but is not limited to the above forms.

[0068] This application can be used in numerous general-purpose or special-purpose computer system environments or configurations. For example: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, distributed computing environments including any of the above systems or devices, and so on. This application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc. that perform specific tasks or implement specific abstract data types. This application can also be practiced in a distributed computing environment where tasks are performed by remote processing devices connected through a communication network. In a distributed computing environment, program modules can be located in local and remote computer storage media including storage devices.

[0069] It should be noted that in each specific embodiment of this application, when it comes to performing relevant processing based on data related to the identity or characteristics of an object, such as object information, object behavior data, object historical data, and object location information, the permission or consent of the object will be obtained first. Moreover, the collection, use, and processing of these data will comply with relevant laws, regulations, and standards. In addition, when the embodiments of this application need to obtain sensitive personal information of an object, the object's separate permission or separate consent will be obtained through methods such as pop-up windows or redirecting to a confirmation page. After clearly obtaining the object's separate permission or separate consent, the necessary object-related data for the normal operation of the embodiments of this application will be obtained.

[0070] Figure 1 is an optional flowchart of the event mining method provided by the embodiments of this application, Figure 1 The method in may include but is not limited to steps S101 to S105:

[0071] Step S101, obtain the target customer complaint text, and the target customer complaint text is sourced from a customer complaint conversation;

[0072] Step S102, perform event extraction on the target customer complaint text based on a pre-trained event generation model to obtain candidate events;

[0073] Step S103, screen out reference events from a preset event library based on the candidate events;

[0074] Step S104, perform similarity detection on the candidate events and the reference events based on a pre-trained first event detection model to obtain an event similarity result;

[0075] Step S105: If the event similarity result indicates that the candidate event is not similar to the reference event, then based on the pre-trained second event detection model and the target customer complaint text, perform event screening on the candidate event to obtain the target mined event.

[0076] Steps S101 to S105 illustrated in the embodiments of the present application, by obtaining the target customer complaint text, where the target customer complaint text is from a customer complaint conversation; based on the pre-trained event generation model, perform event extraction on the target customer complaint text to obtain candidate events, which can initially obtain events from the target customer complaint text. Compared with manual extraction, using the event generation model can be more efficient. At the same time, extracting candidate events from the target customer complaint text provides a structured input for subsequent steps, which is beneficial to improving the matching efficiency and accuracy. Further, based on the candidate events, screen out the reference events from the preset event library. Compared with traversing the event library to match each event in the event library with the candidate events one by one, the pre-screening operation can filter out irrelevant events in the event library, reducing the matching calculation amount and improving the accuracy of event mining. Further, based on the pre-trained first event detection model, perform similarity detection on the candidate event and the reference event to obtain the event similarity result, which can identify whether the candidate event already exists in the event library, thus avoiding redundant information in the event library. Further, if the event similarity result indicates that the candidate event is not similar to the reference event, then based on the pre-trained second event detection model and the target customer complaint text, perform event screening on the candidate event to obtain the target mined event. By performing semantic screening on the dissimilar candidate events based on the target customer complaint text, the accuracy of candidate event generation is improved, thereby improving the accuracy of event mining.

[0077] In step S101 of some embodiments, the customer complaint conversation is a conversation between a customer and a robot customer service or a human customer service. The conversation content includes the questions raised by the customer and the corresponding solutions given by the customer service. The customer complaint conversation can be in text form, audio form, or a form combining text and audio, which is not limited here. The target customer complaint text is from the customer complaint conversation, and the target customer complaint text is the text data to be subjected to event extraction. When the customer complaint conversation is in audio form, the target customer complaint text is obtained through the following steps:

[0078] Obtain the audio data of the customer complaint conversation;

[0079] Perform conversion processing on the audio data of the customer complaint conversation to obtain the original customer complaint text;

[0080] Perform correction processing on the original customer complaint text to obtain the intermediate customer complaint text;

[0081] Perform replacement processing on the intermediate customer complaint text through a preset term library to obtain the target customer complaint text.

[0082] Among them, the audio data of the customer complaint conversation can be obtained through recording during the voice conversation or from a third-party platform, without limitation here. The original customer complaint text is the preliminary text obtained by converting the audio data into text form. The intermediate customer complaint text is the intermediate text obtained after the original customer complaint text has been corrected. The target customer complaint text is the target text obtained after the intermediate customer complaint text has been replaced. The preset term library stores professional terms and non-professional term descriptions corresponding to the professional terms. The preset term library can be constructed by referring to industry-related standards or based on historical data, without limitation here. Moreover, the preset term library can be classified hierarchically according to specific application scenarios, which can shorten the search time and improve the replacement efficiency. The conversion process can be achieved through speech recognition technology, and the speech recognition technology can be at least one of DNN, HMM, RNN, and CNN. The correction process includes spelling correction, grammar correction, and context correction, etc. Spelling correction can be achieved through a lexical model or by calculating the edit distance. Grammar correction can be achieved by judging whether the grammar is reasonable through grammar rules and then correcting the grammar, or by using a hidden Markov model. Context correction can be achieved through an N-gram language model, by keyword matching, or by constructing a knowledge graph.

[0083] In a specific implementation, first, the audio data of the customer complaint conversation is obtained. Then, the audio data of the customer complaint conversation is converted through speech recognition technology to obtain the original customer complaint text. Next, spelling correction, grammar correction, and context correction are performed on the original customer complaint text to obtain the intermediate customer complaint text. Finally, the intermediate customer complaint text is segmented to obtain the keywords corresponding to the intermediate customer complaint text. The non-professional term descriptions are matched from the preset term library using the keywords corresponding to the intermediate customer complaint text, and then the professional terms corresponding to the non-professional term descriptions are found. The professional terms are replaced with the corresponding keywords in the intermediate customer complaint text to obtain the target customer complaint text.

[0084] Before step S102 in some embodiments, the event mining method further includes pre-training an event generation model, which is used to generate events based on the target customer complaint text. The event generation model is constructed in the following manner:

[0085] Generate a prompt text according to the event mining scenario to obtain the prompt text;

[0086] Based on the prompt text and a predetermined large language model, a model is constructed to obtain the event generation model.

[0087] Among them, the event mining scenario is the specific scenario of event mining, such as the customer complaint scenario. The prompt text is used to enable a pre-trained large language model to generate events for the scenario in the prompt text. In the customer complaint event mining scenario, usually the customer raises a question and the customer service provides a solution to the question. Therefore, in the customer complaint event mining scenario, the prompt text can be "I will provide a piece of customer complaint text. Please output events from the customer complaint text according to the requirements. The requirements are to extract the fragments of the key questions raised by the customer and the corresponding answers from the original text, without adding or changing a single word, and output 1 to 3 fragments, each fragment not exceeding 40 characters in length." After obtaining the prompt text, the prompt text is used as the Prompt and input into the large language model to obtain an event generation model. The large language model (LLM) can be GPT or Bert.

[0088] After the event generation model is constructed, it is necessary to train the event generation model. The training process of the event generation model includes: obtaining sample customer complaint texts and the ideal events corresponding to the sample customer complaint texts, inputting the sample customer complaint texts and events into the event generation model, so that the event generation model learns how to output the ideal events corresponding to the sample customer complaint texts based on the sample customer complaint texts, and optimizing the event generation model according to the difference between the real events generated by the event generation model based on the sample customer complaint texts and the ideal events until the difference is less than the set threshold, and stopping the optimization of the event generation model to obtain an event generation model that meets the requirements.

[0089] In step S102 of some embodiments, the candidate events are the events extracted by the event generation model from the target customer complaint text. In a specific implementation, the target customer complaint text is input into the event generation model, and the event generation model outputs the candidate events. First, during the learning process of the end-to-end generation model, since small batches of data are used for training, the learning ability of the model is restricted by the data scale and may not fully capture the huge and complex language patterns and contexts. In contrast, the LLM learns language representations through large-scale pre-training data, can better grasp rich language information, and improves the language understanding ability of the generation model. Second, the end-to-end generation model has limitations in processing text length. During training, it uses a fixed-length input sequence, usually limited to within 512 tokens. However, in many actual scenarios, the text length often exceeds this limit, resulting in the model being unable to fully capture the semantic relationships and information in the long text. On the contrary, the LLM uses a larger text window during pre-training, enabling the model to better understand and process longer text fragments and improving the comprehensive understanding ability of the context. Therefore, introducing the LLM technology for event generation achieves better event generation effects compared to the traditional end-to-end generation model.

[0090] Please refer to Figure 2 In some embodiments, step S103 may include but is not limited to steps S201 to S202:

[0091] Step S201, determining keywords based on a candidate event to obtain target keywords;

[0092] Step S202, filtering events from a preset event library based on the target keywords to obtain reference events.

[0093] In step S201 of some embodiments, the target keywords are keywords in the candidate event that can significantly represent the subject and content of the candidate event. In a specific implementation, first, text preprocessing is performed on the candidate event to obtain a set of words. The preprocessing includes removing stop words, punctuation marks, and performing word segmentation, etc. Then, the term frequency and importance of each word in the set of words are calculated by TF-IDF to obtain the TF-IDF value of each word. Then, keywords are extracted according to the TF-IDF values, and the words with TF-IDF values higher than a set threshold are selected as keywords. Finally, the part of speech can also be filtered according to specific task requirements, that is, the obtained keywords are filtered to filter out some irrelevant parts of speech, such as retaining nouns or verbs, etc., and finally the target keywords are obtained.

[0094] In step S202 of some embodiments, the preset event library is a database for storing events that is continuously updated. Specifically, what is stored are events, and the events include customer complaint problems and corresponding solutions to the customer complaint problems. The reference events are the events in the preset event library that are most similar to the candidate event, and there can be one or more. In a specific implementation, the text data in the preset event library can be inverted index established. Based on the target keywords, a query is made on the inverted index, and the event library returns a list of events containing the target keywords. Then, the returned list of events is sorted, and the most relevant events are ranked at the front. At the same time, irrelevant events can also be filtered according to requirements, or only a preset number of top-ranked events can be obtained. Finally, according to the sorting and filtering results, the final list of events, that is, the reference events, is presented.

[0095] Through the above steps S201 to S202, it is possible to quickly locate the reference events related to the candidate event from the preset event library based on the target keywords, which is beneficial to improving the matching quality and efficiency of the event mining process, thereby improving the accuracy of event mining.

[0096] Before step S104 of some embodiments, the event mining method further includes pre-training a first event detection model for detecting the similarity between a candidate event and a reference event. The first event detection model can also be referred to as an event similarity detection model. Specifically, the first event detection model can be constructed based on the BM25 model, the Word2Vec model, or the BERT model. Taking the BERT model as an example, first, the BERT model is used as the basic model, and the BERT model is fine-tuned using a small amount of pre-constructed original datasets with annotations. Then, data augmentation is performed on the original datasets, including operations such as synonym replacement, word insertion, word deletion, and sentence recombination on the original datasets to generate more augmented datasets. The original datasets and the augmented datasets are divided into training sets and validation sets. The fine-tuned model is re-trained using the training sets, and the performance of the trained model is evaluated using the validation sets. Finally, the first event detection model is constructed.

[0097] Please refer to Figure 3 , Figure 3 is the principle flowchart of the first event detection model provided by the embodiments of the present application. The candidate event and the reference event are respectively vectorized to obtain a candidate event vector and a reference event vector. Then, the event similarity between the candidate event vector and the reference vector is obtained, and the event similarity result is obtained by comparing the event similarity with a preset similarity threshold. The first event detection model includes a vectorization layer, a similarity calculation layer, and a comparison layer. Among them, the vectorization layer is used to vectorize text information. The input of the vectorization layer is an event, and the output is the vectorized representation corresponding to the event, that is, a vector. The similarity calculation layer is used to calculate the similarity between two vectors. The input of the similarity calculation layer is the candidate event vector and the reference event vector, and the output is the event similarity. Among them, the similarity calculation can be realized by calculating the cosine similarity. The comparison layer is used to compare the event similarity with the preset similarity threshold. The input of the comparison layer is the event similarity, and the output is the event similarity result. Substantially, the candidate event and the reference event with a similarity greater than or equal to the preset similarity threshold are determined to be event-similar.

[0098] After the first event detection model is constructed, the training process of the first event detection model includes: obtaining events from a preset event library, randomly combining the events to construct positive and negative samples. The positive samples are similar event pairs, and the negative samples are dissimilar event pairs. Then, the first event detection model is optimized and trained based on the positive and negative samples.

[0099] In step S104 of some embodiments, the event similarity result includes indicating that the candidate event is similar to the reference event or the candidate event is not similar to the reference event. The event similarity result is obtained by performing similarity detection on the candidate event and the reference event based on the first event detection model. In a specific implementation, if the number of reference events is one, the candidate event is vectorized through the vectorization layer of the first event detection model to obtain a candidate event vector. At the same time, the reference event is vectorized through the vectorization layer of the first event detection model to obtain a reference event vector. After obtaining the candidate event vector and the reference event vector, the similarity between the candidate event vector and the reference event vector is calculated based on the similarity calculation layer of the first event detection model to obtain the event similarity. Then, the preset similarity threshold and the event similarity are compared through the comparison layer of the first event detection model. When the event similarity is greater than or equal to the preset similarity threshold, the event similarity result is determined to be similar; when the event similarity is less than the preset similarity threshold, the event similarity result is determined to be dissimilar. Among them, the preset similarity threshold can be set according to actual needs. If the number of reference events is multiple, for each reference event, the similarity between the reference event vector corresponding to the reference event and the candidate event vector is calculated to obtain the event similarity corresponding to each reference event, and the similarity with the largest event similarity value is selected for comparison with the preset similarity threshold to obtain the event similarity result.

[0100] Please refer to Figure 4 , after step S104 of some embodiments, the event mining method may further include, but is not limited to, steps S401 to S402:

[0101] Step S401, if the event similarity result indicates that the candidate event is similar to the reference event, the candidate event and the reference event are merged to obtain a merged event;

[0102] Step S402, the merged event replaces the reference event.

[0103] In step S401 of some embodiments, if the event similarity result indicates that the candidate event is similar to the reference event, it means that the candidate event and the reference event highly match in terms of event attributes. Then, the candidate event and the reference event are merged to obtain a merged event. Among them, the event attributes include the subject, time, location, details, etc. The merging process refers to integrating all event attributes of the candidate event and the reference event into one event description, eliminating duplicate information, and forming a new event.

[0104] Please refer to Figure 5 , the process of merging the candidate event and the reference event to obtain a merged event may include, but is not limited to, steps S501 to S506:

[0105] Step S501: Extract keywords from the candidate event to obtain candidate keywords;

[0106] Step S502: Extract keywords from the reference event to obtain reference keywords;

[0107] Step S503: Calculate the distance between the candidate keywords and the reference keywords to obtain the keyword feature distance between the candidate keywords and the reference keywords;

[0108] Step S504: Construct an event graph based on the candidate keywords, the reference keywords, and the keyword feature distance; wherein, the candidate keywords and the reference keywords serve as the nodes of the event graph, and the keyword feature distance serves as the edge of the event graph;

[0109] Step S505: Perform clustering calculation on the nodes in the event graph to obtain at least one keyword node group;

[0110] Step S506: Reconstruct sentences for each keyword node group based on the reference event and the candidate event to obtain a merged event.

[0111] In step S501 of some embodiments, the candidate keywords are the keywords extracted from the candidate event. Since the specific implementation process of step S501 is similar to that of step S201, it will not be elaborated here.

[0112] In step S502 of some embodiments, the reference keywords are the keywords extracted from the reference event. Since the specific implementation process of step S502 is similar to that of step S201, it will not be elaborated here.

[0113] In step S503 of some embodiments, the keyword feature distance between the candidate keywords and the reference keywords represents the semantic similarity. The closer the distance, the more similar; the farther the distance, the less similar. Among them, the distance calculation can be achieved by calculating the cosine similarity, or by calculating the Euclidean distance, or by calculating the edit distance, which is not limited here.

[0114] In step S504 of some embodiments, the event graph is an undirected graph. The nodes of the event graph are the candidate keywords and the reference keywords, and the edges of the event graph are the keyword feature distances between the candidate keywords and the reference keywords. In specific implementation, the candidate keywords and the reference keywords are used as nodes, and then edges are added to the corresponding nodes based on the keyword feature distance to obtain the event graph. By constructing the event graph, the relationship between the candidate keywords and the reference keywords can be intuitively presented, making the complex association relationship easier to understand.

[0115] In step S505 of some embodiments, the number of keyword node groups is at least one. In each keyword node group, it can be a combination of a candidate keyword and a reference keyword, or only one candidate keyword, or only one reference keyword. Clustering calculation can be implemented by using a clustering algorithm, such as the K-means algorithm. Through clustering calculation, keywords can be grouped to reduce information complexity and redundant information.

[0116] In step S506 of some embodiments, sentence recombination refers to the recombination of candidate events and reference events to generate a merged event that includes both candidate events and reference events. During the process of sentence recombination, the aim is to eliminate redundancy, retain important information, and ensure that the generated sentence still accurately describes the same event or topic semantically. In a specific implementation, a representative keyword is selected from each keyword node group. That is, if there is only one node in the keyword node group, the keyword corresponding to that node is used as the representative keyword. If there are two nodes in the keyword node group, a keyword corresponding to one of the two nodes is randomly selected as the representative keyword. Based on the representative keyword, the text fragment where the representative keyword is located is extracted from the reference event or candidate event, and all the text fragments are recombined into a sentence to obtain the merged event. This step helps to eliminate redundant information, refine key content, improve the expressiveness and accuracy of the generated merged event, and provide a more comprehensive event description.

[0117] Through the above steps S501 to S506, technical means such as keyword extraction, keyword feature distance calculation, and event graph clustering can be comprehensively utilized. By deeply exploring the relationships between event elements, the merging of candidate events and reference events is achieved, improving the accuracy, comprehensiveness, and conciseness of event descriptions.

[0118] In step S402 of some embodiments, after obtaining the merged event, the merged event is used to replace the reference event in the preset event library. It should be noted that if multiple reference events are filtered out based on the candidate event, during the merging process, the candidate event is merged with the reference event that has the highest event similarity with the candidate event to obtain the merged event, and the merged event is used to replace the reference event in the preset event library that has the highest event similarity with the candidate event and the reference event. Suppose the reference event is "The express delivery was signed for privately and the package was lost, please claim compensation", and the candidate event is "The express delivery was signed for privately and the package was lost, please file a complaint". The merged event can be "The express delivery was signed for privately and the package was lost, please claim compensation or file a complaint".

[0119] Through the above steps S401 to S402, multiple records describing the same event can be merged into one through similarity calculation, avoiding duplicate information in the preset event library, thereby reducing redundancy. At the same time, the preset event library is updated and improved, which is beneficial to improving the accuracy of event recognition and avoiding multiple identifications of the same event.

[0120] Before step S105 in some embodiments, the event mining method further includes pre-training a second event detection model, which is used to detect the semantic validity of candidate events. The second event detection model can also be called an event validity detection model. Specifically, the second event detection model can be constructed based on the BERT entailment model.

[0121] The second event detection model includes a preprocessing layer, a Transformer layer, a pooling layer, and a fully connected layer. Among them, the preprocessing layer is used to perform preliminary processing on candidate events. The preliminary processing includes operations such as removing stop words, punctuation marks, and word segmentation. The input of the preprocessing layer is candidate events, and the output is a sequence of word vectors. The Transformer layer is used to encode the sequence of word vectors, capturing long-distance dependencies in the text sequence through the self-attention mechanism. The input of the Transformer layer is the sequence of word vectors, and the output is a sequence of hidden states. The pooling layer is used to extract semantic information vectors from the sequence of hidden states. The input of the pooling layer is the sequence of hidden states, and the output is a semantic information vector. The fully connected layer is used to judge whether the information of the first semantic information vector is contained in the information of the second semantic information vector. The input is the first semantic information vector and the second semantic information vector, and the output is a probability distribution result.

[0122] After constructing the second event detection model, the training process of the second event detection model includes: all events in the preset event library are valid. Randomly select valid events from the preset event library, randomly truncate the selected valid events, and obtain a part of the text information in the valid events as negative samples, ensuring that the negative samples still have the context of valid events, but the overall information of the negative samples is not sufficient to represent a complete valid event. At the same time, obtain the complete text information of the valid events as positive samples. Combine the obtained negative samples and positive samples to form a training set, and use the training set to optimize and train the second event detection model.

[0123] It should be noted that the first event detection model only needs to support similarity judgment through a small amount of artificially constructed tags and data during cold start, that is, to support data migration applications of the basic model in the field of event similarity judgment. At the same time, the second event detection model also only needs to support semantic judgment through a small amount of artificially constructed tags and data during cold start, that is, to support data migration applications of the basic model in the field of event semantic judgment. Then, the model after cold start is trained based on open-source data. Therefore, only a small amount of manually annotated data is required in the embodiments of the present application.

[0124] In step S105 of some embodiments, the target mining event is a new event not stored in the preset event library. If the event similarity result indicates that the candidate event and the reference event are not similar, then the candidate event is screened based on the pre-trained second event detection model and the target customer complaint text to obtain the target mining event.

[0125] Please refer to Figure 6 , the process of screening the candidate event based on the pre-trained second event detection model and the target customer complaint text to obtain the target mining event may include, but is not limited to, steps S601 to S606:

[0126] Step S601, perform word segmentation and embedding processing on the candidate event based on the preprocessing layer to obtain a word vector sequence;

[0127] Step S602, perform encoding processing on the word vector sequence based on the Transformer layer to obtain a hidden state sequence;

[0128] Step S603, perform information extraction on the hidden state sequence based on the pooling layer to obtain a first semantic information vector;

[0129] Step S604, obtain the semantic information vector of the target customer complaint text to obtain a second semantic information vector;

[0130] Step S605, perform classification processing on the first semantic information vector based on the fully connected layer and the second semantic information vector to obtain a probability distribution result;

[0131] Step S606, screen the candidate event based on the probability distribution result to obtain the target mining event.

[0132] In step S601 of some embodiments, the word vector sequence is obtained by word segmentation and embedding processing, which maps each word in the text of the candidate event to the corresponding word vector and forms a sequence according to the order of the words in the candidate event.

[0133] In step S602 of some embodiments, the hidden state sequence is a sequence of context representations obtained by encoding each word through the Transformer layer. This sequence is sorted according to the order of words in the candidate event and corresponds one-to-one with the elements in the word vector sequence. Each element in the hidden state sequence corresponds to the hidden state of a word, and the hidden state contains the feature information learned at the time step or sample.

[0134] In step S603 of some embodiments, the first semantic information vector is a vector representing the overall semantic information of the candidate event. The pooling layer can be average pooling or max pooling to extract the important information of the entire hidden state sequence and obtain the first semantic information vector.

[0135] In step S604 of some embodiments, the second semantic information vector is a vector representing the overall semantic information of the target customer complaint text. The acquisition method of the second semantic information vector is similar to that of the first semantic information vector and will not be elaborated here.

[0136] In step S605 of some embodiments, the probability distribution result includes that the second semantic information vector contains the first semantic information vector or the second semantic information vector does not contain the first semantic information vector. Based on the fully connected layer and the second semantic information vector, a probability distribution is generated for the first semantic information vector, and the category with the highest probability is used as the probability distribution result. The categories refer to containment and non - containment.

[0137] In step S606 of some embodiments, if the probability distribution result is that the second semantic information vector contains the first semantic information vector, it indicates that the semantics of the candidate event is valid. Then, the candidate event is used as the target mining event and the target mining event is stored in the preset event library. If the probability distribution result is that the second semantic information vector does not contain the first semantic information vector, it indicates that the semantics of the candidate event is invalid, and then the candidate event is deleted.

[0138] Through the above steps S601 to S606, it is possible to avoid obtaining untrue or invalid events due to the certain degree of uncertainty and incomplete control in event generation by the LLM, thereby improving the reliability of event generation and the quality of events.

[0139] Please refer to Figure 7 , Figure 7 which is a schematic diagram of the specific implementation process of the event mining method provided by the embodiments of the present application. The specific implementation process of extracting candidate events from the target customer complaint text through the event generation model is similar to the specific implementation process of the above step S102. For the sake of brevity, it will not be elaborated here. The specific implementation process of performing similarity detection on the candidate events through the first event detection model is similar to the specific implementation process of the above step S103. For the sake of brevity, it will not be elaborated here.

[0140] If the event similarity detection detects similarity, the specific implementation process of event merging is similar to the specific implementation processes of steps S401 to S402 above. To save space, it will not be elaborated here. After event merging, the merged event is stored in the event library to replace the reference event corresponding to the merged event. If the event similarity detection detects dissimilarity, the specific implementation process of effectively detecting the candidate event through the second event detection model is similar to the specific implementation processes of steps S601 to S606 above. To save space, it will not be elaborated here. If the candidate event is detected to be valid, the candidate event is stored in the event library. If the candidate event is detected to be invalid, the candidate event is deleted. The event library is also used to construct positive and negative samples based on the data in the event library for random combination to optimize the first event detection model, and the event library is also used to construct positive and negative samples based on the data in the event library for random truncation to optimize the second event detection model. This method realizes a system closed-loop and can reverse-optimize each model in the system link based on the updated event library. At the same time, there is a bottleneck in the inference speed of the LLM, and the generation ability of the traditional BERT model is not good. However, the present application can improve the closed-loop advantage by combining the LLM with a small model of the traditional BERT.

[0141] Please refer to Figure 8 For this, the embodiment of the present application also provides an event mining system, which can implement the above event mining method. The system includes:

[0142] A data acquisition module 801, configured to acquire a target customer complaint text, and the target customer complaint text is from a customer complaint conversation;

[0143] A candidate event acquisition module 802, configured to perform event extraction on the target customer complaint text based on a pre-trained event generation model to obtain candidate events;

[0144] A reference event acquisition module 803, configured to screen out reference events from a preset event library based on the candidate events;

[0145] A first detection module 804, configured to perform similarity detection on the candidate event and the reference event based on a pre-trained first event detection model to obtain an event similarity result;

[0146] A second detection module 805, configured to, if the event similarity result indicates that the candidate event and the reference event are dissimilar, perform event screening on the candidate event based on a pre-trained second event detection model and the target customer complaint text to obtain a target mining event.

[0147] The specific implementation manner of this event mining system is basically the same as the specific embodiment of the above event mining method, and will not be elaborated here.

[0148] The embodiments of the present application also provide an electronic device, which includes a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, the above-mentioned event mining method is implemented. The electronic device can be any intelligent terminal including a tablet computer, an in-vehicle computer, etc.

[0149] Please refer to Figure 9 , Figure 9 which schematically shows the hardware structure of an electronic device according to another embodiment. The electronic device includes:

[0150] A processor 901, which can be implemented by using a general-purpose CPU (Central Processing Unit), a microprocessor, an application-specific integrated circuit (ASIC), or one or more integrated circuits, etc., and is used to execute relevant programs to implement the technical solutions provided by the embodiments of the present application;

[0151] A memory 902, which can be implemented in the form of a read-only memory (ROM), a static storage device, a dynamic storage device, or a random access memory (RAM), etc. The memory 902 can store an operating system and other application programs. When implementing the technical solutions provided by the embodiments of this specification through software or firmware, the relevant program codes are stored in the memory 902, and the processor 901 is called to execute the event mining method of the embodiments of the present application;

[0152] An input / output interface 903, which is used to implement information input and output;

[0153] A communication interface 904, which is used to implement communication interaction between this device and other devices, and can implement communication through a wired method (such as USB, network cable, etc.) or through a wireless method (such as a mobile network, WIFI, Bluetooth, etc.);

[0154] A bus 905, which transmits information between various components of the device (such as the processor 901, the memory 902, the input / output interface 903, and the communication interface 904);

[0155] Among them, the processor 901, the memory 902, the input / output interface 903, and the communication interface 904 are communicatively connected to each other inside the device through the bus 905.

[0156] The embodiments of the present application also provide a computer-readable storage medium, which stores a computer program, and when the computer program is executed by a processor, the above-mentioned event mining method is implemented.

[0157] As a non-transitory computer-readable storage medium, the memory can be used to store non-transitory software programs and non-transitory computer-executable programs. In addition, the memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one magnetic disk storage device, a flash memory device, or other non-transitory solid-state storage devices. In some embodiments, the memory optionally includes a memory remotely disposed relative to the processor, and these remote memories can be connected to the processor through a network. Examples of the above networks include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0158] The event mining method, event mining system, electronic device, and storage medium provided by the embodiments of the present application obtain a target customer complaint text, and the target customer complaint text is derived from a customer complaint conversation; based on a pre-trained event generation model, the target customer complaint text is subjected to event extraction to obtain candidate events, which can initially obtain events from the target customer complaint text. Compared with manual extraction, the event generation model can be more efficient. At the same time, extracting candidate events from the target customer complaint text provides a structured input for subsequent steps, which is beneficial to improving the matching efficiency and accuracy. Further, based on the candidate events, reference events are screened out from a preset event library. Compared with traversing the event library to match the events in the event library with the candidate events one by one, the pre-screening operation can filter out irrelevant events in the event library, reduce the matching calculation amount, and improve the accuracy of event mining. Further, based on a pre-trained first event detection model, the candidate events and the reference events are subjected to similarity detection to obtain an event similarity result, which can identify whether the candidate events already exist in the event library, thereby avoiding redundant information in the event library. Further, if the event similarity result indicates that the candidate events and the reference events are not similar, then based on a pre-trained second event detection model and the target customer complaint text, the candidate events are screened to obtain target mining events. By semantically screening the dissimilar candidate events based on the target customer complaint text, the accuracy of candidate event generation is improved, thereby improving the accuracy of event mining. At the same time, the embodiments of the present application also provide a complete event generation and screening strategy.

[0159] The embodiments described in the embodiments of the present application are for more clearly illustrating the technical solutions of the embodiments of the present application, and do not constitute a limitation on the technical solutions provided by the embodiments of the present application. Those skilled in the art will know that with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by the embodiments of the present application are equally applicable to similar technical problems.

[0160] Those skilled in the art can understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of the present application, and may include more or fewer steps than shown in the figures, or combine certain steps, or different steps.

[0161] The system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separated, that is, they may be located in one place or distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0162] Those of ordinary skill in the art can understand that all or some of the steps in the methods disclosed above, and the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, and their appropriate combinations.

[0163] The terms "first", "second", "third", "fourth", etc. (if any) in the specification of this application and the above-mentioned drawings are used to distinguish similar objects and do not necessarily describe a specific order or sequence. It should be understood that the data used in this way can be interchanged under appropriate circumstances so that the embodiments of this application described here can be implemented in an order other than those illustrated or described here. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units does not necessarily have to be limited to those steps or units clearly listed, but may include other steps or units not clearly listed or inherent to these processes, methods, products, or devices.

[0164] It should be understood that in this application, "at least one (item)" means one or more, and "a plurality" means two or more. "And / or" is used to describe the association relationship of associated objects and indicates that three relationships can exist. For example, "A and / or B" can mean: only A exists, only B exists, and both A and B exist at the same time. Among them, A and B can be singular or plural. The character " / " generally means that the associated objects before and after are in an "or" relationship. "At least one (one) of the following" or its similar expression refers to any combination of these items, including any combination of single item (one) or plural items (ones). For example, at least one (one) of a, b, or c can mean: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, c can be single or multiple.

[0165] In several embodiments provided by the present application, it should be understood that the disclosed systems and methods can be implemented in other ways. For example, the system embodiments described above are merely illustrative. For example, the division of the above units is only a logical function division. In actual implementation, there may be other division methods. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. The displayed or discussed couplings or direct couplings or communication connections to each other can be through some interfaces. The indirect couplings or communication connections of systems or units can be in electrical, mechanical or other forms.

[0166] The units described above as separate components may or may not be physically separated. The components displayed as units may or may not be physical units, that is, they may be located in one place, or may be distributed to multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0167] In addition, in each embodiment of the present application, the functional units can be integrated in a processing unit, or each unit can exist physically alone, or two or more units can be integrated in one unit. The above integrated units can be implemented in the form of hardware or in the form of software functional units.

[0168] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present application, in essence, or the part that contributes to the prior art, or all or 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 multiple instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods in each embodiment of the present application. The foregoing storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memories (ROM), random access memories (RAM), magnetic disks, or optical discs that can store programs.

[0169] The preferred embodiments of the embodiments of the present application have been described above with reference to the accompanying drawings, but this does not limit the scope of rights of the embodiments of the present application. Any modifications, equivalent replacements, and improvements made by those skilled in the art without departing from the scope and essence of the embodiments of the present application shall be within the scope of rights of the embodiments of the present application.

Claims

1. An event mining method, characterized in that, The method includes: Obtaining a target customer complaint text, where the target customer complaint text is from a customer complaint conversation; Performing event extraction on the target customer complaint text based on a pre-trained event generation model to obtain candidate events; Screening out reference events from a preset event library based on the candidate events; Performing similarity detection on the candidate events and the reference events based on a pre-trained first event detection model to obtain an event similarity result; If the event similarity result indicates that the candidate event is not similar to the reference event, then performing event screening on the candidate event based on a pre-trained second event detection model and the target customer complaint text to obtain a target mining event.

2. The method according to claim 1, wherein After performing similarity detection on the candidate events and the reference events based on a pre-trained first event detection model to obtain an event similarity result, the method further includes: If the event similarity result indicates that the candidate event is similar to the reference event, then merging the candidate event and the reference event to obtain a merged event; Replacing the reference event with the merged event.

3. The method according to claim 2, wherein The merging the candidate event and the reference event to obtain a merged event includes: Extracting keywords from the candidate event to obtain candidate keywords; Extracting keywords from the reference event to obtain reference keywords; Calculating the distance between the candidate keywords and the reference keywords to obtain a keyword feature distance between the candidate keywords and the reference keywords; Constructing an event graph based on the candidate keywords, the reference keywords, and the keyword feature distance; where the candidate keywords and the reference keywords are used as nodes of the event graph, and the keyword feature distance is used as an edge of the event graph; Performing clustering calculation on the nodes in the event graph to obtain at least one keyword node group; Recombining sentences for each keyword node group based on the reference event and the candidate event to obtain the merged event.

4. The method according to claim 1, wherein The screening out reference events from a preset event library based on the candidate events includes: Determining target keywords based on the candidate event; Screening events from the preset event library based on the target keywords to obtain the reference events.

5. The method according to claim 1, characterized in that, The pre-trained second event detection model includes a preprocessing layer, a Transformer layer, a pooling layer, and a fully connected layer. The performing event screening on the candidate event based on a pre-trained second event detection model and the target customer complaint text to obtain a target mining event includes: Performing word segmentation and embedding processing on the candidate event based on the preprocessing layer to obtain a word vector sequence; Performing encoding processing on the word vector sequence based on the Transformer layer to obtain a hidden state sequence; Performing information extraction on the hidden state sequence based on the pooling layer to obtain a first semantic information vector; Obtaining a second semantic information vector by acquiring the semantic information vector of the target customer complaint text; Performing classification processing on the first semantic information vector based on the fully connected layer and the second semantic information vector to obtain a probability distribution result; Event screening is performed on the candidate events based on the probability distribution results to obtain the target mining events.

6. The method according to claim 1, wherein The event generation model is constructed in the following manner: Prompt generation is performed according to the event mining scenario to obtain prompt text; Based on the prompt text and a predetermined large language model, model construction is carried out to obtain an event generation model.

7. The method according to any one of claims 1 to 6, characterized in that, The target customer complaint text is obtained in the following manner: Audio data of the customer complaint conversation is obtained; The audio data of the customer complaint conversation is subjected to conversion processing to obtain the original customer complaint text; The original customer complaint text is subjected to correction processing to obtain the intermediate customer complaint text; The intermediate customer complaint text is subjected to replacement processing through a preset term library to obtain the target customer complaint text.

8. An event mining system, characterized in that, The system includes: A data acquisition module for acquiring target customer complaint text, where the target customer complaint text is derived from a customer complaint conversation; A candidate event acquisition module for performing event extraction on the target customer complaint text based on a pre-trained event generation model to obtain candidate events; A reference event acquisition module for screening reference events from a preset event library based on the candidate events; A first detection module for performing similarity detection on the candidate events and the reference events based on a pre-trained first event detection model to obtain an event similarity result; A second detection module for, if the event similarity result indicates that the candidate event is not similar to the reference event, performing event screening on the candidate event based on a pre-trained second event detection model and the target customer complaint text to obtain the target mining event.

9. An electronic device, characterized in that, The electronic device includes a memory and a processor, the memory stores a computer program, and when the processor executes the computer program, the event mining method according to any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, the event mining method according to any one of claims 1 to 7 is implemented.

Images