Premium collection call script extraction and analysis method, apparatus, device, and medium

Abstract

The application provides a premium collection script extraction and analysis method, device, equipment and medium, text data of premium collection telephone recording is obtained, and the text data is classified to screen out conversation text of no renewal willingness category; each sentence in the conversation text is subject classified to analyze theme missing rate and sequence difference between conversation texts of each seat with different success rates; meaningless sentences are deleted, keywords of each sentence are extracted to analyze word difference of keywords in the same theme in conversation texts of each seat with different success rates, and / or conversation texts of each seat with different success rates under each theme are scored to summarize high-score conversation texts. The application can more easily locate excellent scripts and words in the renewal retention business, thereby improving the success rate of seat telephone renewal.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a method, apparatus, equipment and medium for extracting and analyzing premium collection scripts. Background Technology

[0002] Premium collection is a crucial part of the insurance industry's service sector. Communication methods and skills directly impact the success of premium collection, as well as the ability to retain customers for policy renewals and generate long-term revenue for the company.

[0003] Traditional methods for improving sales techniques involve recording and summarizing the experiences and skills of experienced agents, integrating them with business logic, and then disseminating these summaries to other agents through training. However, the difficulty in accurately and concretely recording and summarizing abstract experiences leads to time-consuming processes and potentially inaccurate summaries.

[0004] With the advancement of science and technology, especially the introduction of natural language processing (NLP) algorithms, many practitioners have attempted to use NLP algorithms to solve the problem of discovering and summarizing successful dialogues. The main methods include automatically generating dialogues using algorithms, extracting tags from the entire dialogue content, and classifying and scoring dialogues to determine their success. However, due to the complexity of dialogue content and the limitations of existing technology, these methods all have various application problems, such as long development cycles, difficulty in accurately locating successful dialogues, difficulty in accurately identifying at which stage of communication the agent encountered a problem, and difficulty in detecting differences in dialogue between different agents.

[0005] Therefore, how to automatically identify which communication skills are excelling among top agents, and how to organize and refine successful sales techniques, is crucial to whether the entire customer service agent team can improve the overall success rate of premium collection. This is also a problem that the industry hopes to solve. Summary of the Invention

[0006] In view of the shortcomings of the prior art described above, the purpose of this application is to provide a method, apparatus, equipment and medium for extracting and analyzing premium collection scripts, so as to solve at least one problem in the extraction and sorting of premium collection scripts in the prior art.

[0007] To achieve the above and other related objectives, this application provides a method for extracting and analyzing premium collection scripts. The method includes: acquiring text data from recorded premium collection calls and classifying the text data to filter out dialogue texts belonging to the category of no intention to renew insurance; classifying each sentence in the dialogue texts by topic to analyze the topic missing rate and order differences between dialogue texts of different agents with different success rates; deleting meaningless sentences, extracting keywords from each sentence to analyze the differences in keyword usage in the same topic in dialogue texts of different agents with different success rates, and / or scoring the dialogue texts of different agents with different success rates under each topic to summarize high-scoring dialogue texts.

[0008] In one embodiment of this application, the method for classifying the text data to filter out dialogue texts with no intention to renew insurance includes: segmenting training texts pre-labeled with different classification tags and generating first word vectors using a BERT model; using the first word vectors as input to a TextCNN model to train the TextCNN model; and performing preliminary classification on the text data using the trained TextCNN model to select dialogue texts with the classification tag of no intention to renew insurance.

[0009] In one embodiment of this application, the method for classifying each sentence in the dialogue text into topics includes: segmenting each sentence of a training text pre-labeled with different topic tags and generating a second word vector using a BERT model; inputting the second word vectors into a sentence-level Bi-LSTM model as a group of sentences to obtain a sentence vector for each sentence; inputting each sentence vector and the interlocutor's identity information into the Bi-LSTM model in dialogue order to obtain a conversation vector corresponding to each sentence; converting the conversation vector into a probability that each sentence is the starting sentence of a topic paragraph using a fully connected SoftMax layer; using the probability as input to a CRF layer and optimizing the probability value to obtain a corrected final probability; training the Bi-LSTM model with training text pre-labeled with different topic tags so that the trained Bi-LSTM model can be used to classify each sentence in the dialogue text into topics.

[0010] In one embodiment of this application, the subject classification includes any one or more of the following: an opening statement, a statement of the reason for the call, an inquiry about the reason for not renewing the insurance, a statement of advantages and disadvantages, and a closing statement.

[0011] In one embodiment of this application, the method for deleting meaningless sentences includes: segmenting all dialogue texts into words, converting them into vectors according to the frequency of each word in each sentence using a bag-of-words model; labeling each sentence as meaningful or meaningless; and constructing a binary classification model using a logistic regression algorithm based on the vectors and labeling results of the dialogue texts to classify the sentences of all dialogue texts, thereby excluding sentences labeled as meaningless.

[0012] In one embodiment of this application, the method for scoring the dialogue texts of each agent with different success rates under each topic to summarize high-scoring dialogue texts includes: segmenting each sentence of a training text pre-labeled with different score labels and generating a third word vector using a BERT model; inputting the third word vectors into a TextCNN model as a group for each paragraph to obtain a score for each paragraph; training the TextCNN model with the training text pre-labeled with different score labels so that the trained TextCNN model can score all dialogue texts.

[0013] In one embodiment of this application, the method includes: determining whether each text data item is renewed based on whether the policy is successfully paid in the current period; determining the success rate based on the number of text data items for each agent and the renewal result; analyzing the dialogue text of each agent with different success rates, and selecting at least one dialogue text item with a success rate higher than the average success rate and one dialogue text item with a success rate lower than the average success rate.

[0014] To achieve the above and other related objectives, this application provides a device for extracting and analyzing premium collection scripts. The device includes: an extraction module for acquiring text data from recorded premium collection calls and classifying the text data to filter out dialogue texts belonging to the category of no intention to renew insurance; a classification module for classifying each sentence in the dialogue texts by topic to analyze the topic missing rate and order differences between dialogue texts of different agents with different success rates; and an analysis module for deleting meaningless sentences, extracting keywords from each sentence to analyze the differences in keyword usage in the same topic in the dialogue texts of different agents with different success rates, and / or scoring the dialogue texts of different agents with different success rates under each topic to summarize high-scoring dialogue texts.

[0015] To achieve the above and other related objectives, this application provides a computer device, the device comprising: a memory and a processor; the memory being used to store computer instructions; and the processor executing the computer instructions to implement the method described above.

[0016] To achieve the above and other related objectives, this application provides a computer-readable storage medium storing computer instructions that, when executed, perform the method described above.

[0017] In summary, the method, apparatus, equipment, and medium for extracting and analyzing premium collection scripts provided in this application acquire text data from recorded premium collection calls, classify the text data to filter out dialogue texts belonging to the category of no intention to renew insurance; classify each sentence in the dialogue text by topic to analyze the topic missing rate and order differences between dialogue texts of different agents with different success rates; delete meaningless sentences, extract keywords from each sentence to analyze the wording differences of keywords in the same topic in dialogue texts of different agents with different success rates, and / or score the dialogue texts of different agents with different success rates under each topic to summarize high-scoring dialogue texts.

[0018] It has the following beneficial effects:

[0019] This application makes it easier to identify effective scripts and phrases for policy renewal and retention, thereby increasing the success rate of telephone renewals by agents. Attached Figure Description

[0020] Figure 1 The diagram shows a flowchart of a method for extracting and analyzing premium collection scripts in one embodiment of this application.

[0021] Figure 2 The diagram shows a simplified illustration of a method for extracting and analyzing premium collection scripts in one embodiment of this application.

[0022] Figure 3 The diagram shown is a flowchart illustrating the topic classification of dialogue text in one embodiment of this application.

[0023] Figure 4 The diagram shown is a schematic of a module of a premium collection script extraction and analysis device in one embodiment of this application.

[0024] Figure 5 The diagram shown is a structural schematic of a computer device according to one embodiment of this application. Detailed Implementation

[0025] The following specific examples illustrate the implementation of this application. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. This application can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this application. It should be noted that, unless otherwise specified, the following embodiments and features in the embodiments can be combined with each other.

[0026] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of this application. Although the illustrations only show the components related to this application and are not drawn according to the actual number, shape and size of the components, the form, quantity and proportion of each component can be arbitrarily changed in actual implementation, and the layout of the components may also be more complex.

[0027] As used herein, the singular forms “a,” “an,” and “the” are intended to include the plural forms as well, unless the context indicates otherwise. It should be further understood that the terms “comprising,” “including,” indicate the presence of the stated feature, step, operation, element, component, item, kind, and / or group, but do not preclude the presence, occurrence, or addition of one or more other features, steps, operations, elements, components, items, kinds, and / or groups. The terms “or” and “and / or” as used herein are to be interpreted inclusively, or mean any one or any combination thereof. Thus, “A, B, or C” or “A, B, and / or C” means “any one of: A; B; C; A and B; A and C; B and C; A, B, and C.” Exceptions to this definition occur only when combinations of elements, functions, steps, or operations are inherently mutually exclusive in some way.

[0028] Traditional script extraction and organization mainly relies on the summaries and teachings of agents, but abstract experience is not accurate enough. Using natural language processing algorithms to extract scripts cannot accurately identify excellent scripts, pinpoint where communication problems occurred, or discover differences in dialogue between different agents.

[0029] To address the aforementioned issues, this application provides a method, apparatus, equipment, and medium for extracting and analyzing premium collection scripts. Specifically, this application targets premium collection and renewal business. By extracting scripts and classifying them, extracting keywords, or scoring them, it comprehensively analyzes the differences between conversations from different agents to identify shortcomings or communication errors. It also compiles excellent scripts, continuously improving and expanding the script library. Simultaneously, it trains agents with low success rates, helping them identify deficiencies in their scripts on various topics, thereby improving renewal success rates.

[0030] like Figure 1 The figure shows a flowchart illustrating a method for extracting and analyzing premium collection scripts according to an embodiment of this application. As shown, the method includes:

[0031] Step S101: Obtain the text data of the premium collection call recording, and classify the text data to filter out the dialogue text of the category of no intention to renew the policy.

[0032] In short, the process involves retrieving audio data from the database of premium payment reminder calls and then converting the audio dialogue into text data. Additionally, the database can be used to retrieve data such as policy numbers, renewal queue numbers, and renewal status (whether renewal was successful). The results of whether the policy was successfully paid in the current period can be used as a reference for summarizing subsequent communication scripts.

[0033] For ease of understanding, the following steps can also be referenced as follows: Figure 2 The flowchart of this method is shown below.

[0034] In one embodiment of this application, the method for classifying the text data to filter out dialogue texts belonging to the category of "no intention to renew insurance" includes:

[0035] A. Segment the training text pre-labeled with different classification tags and use the BERT model to generate the first word vector;

[0036] B. Use the first word vector as input to the TextCNN model to train the TextCNN model;

[0037] C. The text data is initially screened and classified using the trained TextCNN model to select dialogue texts with the classification label of "no intention to renew insurance".

[0038] In short, the first step is to categorize the text data based on the content of the conversation. Through observation and understanding, it was found that actual collection calls mainly fall into the following five categories: "The customer informs that they have already paid," "The call is too short or the customer hangs up directly," "The customer indicates that they intended to renew their policy but were unable to pay on time for some reason," "The agent was unable to contact the customer, so they called the salesperson to ask the salesperson to try to contact them on their behalf," and "The customer has no intention of renewing their policy."

[0039] Then, the text data is classified, and the text is labeled as belonging to one of the above five categories. A text classification model is built using BERT+TextCNN to perform initial classification of the dialogue content.

[0040] BERT (Bidirectional Encoder Representations from Transformers) is a pre-trained model proposed by Google in 2018, which is the encoder of bidirectional Transformers.

[0041] The TextCNN model is a model proposed by Yoon Kim in his paper "Convolutional Naural Networks for Sentence Classification" that uses convolutional neural networks to handle NLP problems. Compared with traditional NLP models such as RNN / LSTM, CNN can extract important features more efficiently.

[0042] It should be noted that, with the aim of improving the success rate of collection by agents and summarizing successful sales techniques, the main task is to retain customers who initially had no intention of renewing their policies and to subsequently discover excellent sales techniques. Therefore, only conversations with "no intention of renewing" were selected.

[0043] The Bert+TextCNN model is used to construct a text classification model. First, the labeled dialogue text data is segmented into words. Then, the Bert model is used to generate word vectors. The word vectors are then used as input to the TextCNN model to train the model. Finally, the labeled text data is used to train the model. The trained model is then used to perform preliminary classification on all the text data to be analyzed, and dialogues with the classification result of "no intention to renew insurance" are selected.

[0044] Step S102: Classify each sentence in the dialogue text by topic to analyze the topic missing rate and order differences between dialogue texts of different agents with different success rates.

[0045] Preferably, the topic classification includes, but is not limited to, several topics such as: opening remarks, stating the reason for the call, inquiring about the reasons for not renewing the policy, stating the advantages and disadvantages, and closing remarks. The main purpose is to divide the dialogue text of "no intention to renew" into several topic paragraphs. This allows for comparison of the differences in the scripts used by agents with higher success rates for renewing "no intention to renew" policies and agents with lower success rates, based on the final renewal results obtained from the database.

[0046] For example: a) Compare whether there are differences in the missing rates of each topic paragraph in the dialogue texts of each agent. For example, what percentage of the dialogue texts of each agent are missing the topic paragraph on "statement of advantages and disadvantages"?

[0047] b) Compare whether the main triggering order of each topic in the dialogue texts of each agent is different. For example, the main triggering order of Agent A is "opening remarks" -> "state the reason for the call" -> "inquire about the reason for not renewing the insurance" -> "state the advantages and disadvantages" -> "closing remarks", while the main triggering order of Agent B is "opening remarks" -> "state the reason for the call" -> "state the advantages and disadvantages" -> "inquire about the reason for not renewing the insurance" -> "closing remarks".

[0048] In some examples, by statistically analyzing the missing topics and the triggering order, we can discover the impact of each topic on the success rate, or the impact of the topic order on the success rate. This can not only standardize the required topics and order of the script, but also identify the key topics that affect the success rate for subsequent methods.

[0049] In one embodiment of this application, the method for classifying each sentence in the dialogue text by topic includes:

[0050] A. Segment each sentence of the training text that has been pre-labeled with different topic tags and use the BERT model to generate a second word vector;

[0051] B. Input the second word vector into the sentence-level Bi-LSTM model as a group for each sentence to obtain the sentence vector for each sentence;

[0052] C. Input each sentence vector and the interlocutor's identity information into the Bi-LSTM model in the order of the dialogue to obtain the conversation vector corresponding to each sentence;

[0053] D. Use a fully connected SoftMax layer to convert the session vector into the probability that each sentence is the starting sentence of a topic paragraph;

[0054] E. Use the probability as input to the CRF layer and optimize the probability value to obtain the corrected final probability;

[0055] F. Train the Bi-LSTM model with training texts pre-labeled with different topic tags, so that the trained Bi-LSTM model can be used to classify each sentence in the dialogue text into a topic.

[0056] like Figure 3 As shown, in simple terms, the specific steps are as follows:

[0057] a) Segment each sentence of the annotated dialogue text and generate word vectors using the BERT model.

[0058] b) Input the word vectors into the sentence-level Bi-LSTM model as a group for each sentence to obtain the sentence vector for each sentence.

[0059] BiLSTM is an abbreviation for Bi-directional Long Short-Term Memory, which is composed of a forward LSTM and a backward LSTM. Both are commonly used in natural language processing tasks to model contextual information.

[0060] c) Input the sentence vector of each sentence and the speaker identity information into the dialogue-level Bi-LSTM model in the order of the conversation, and obtain a vector representation for each sentence in the conversation.

[0061] d) Use a fully connected SoftMax layer to convert the vector obtained in the previous step into the probability that each sentence is the starting sentence of a certain topic paragraph.

[0062] e) Use the probability obtained in the previous step as the input to the CRF layer to optimize the probability value and obtain the corrected final probability.

[0063] f) Train the model using labeled text data, and use the trained model to classify the sentences of all texts to be analyzed into topics.

[0064] Step S103: Delete meaningless sentences, extract keywords from each sentence to analyze the differences in keyword usage in the same topic among the dialogue texts of each agent with different success rates, and / or score the dialogue texts of each agent with different success rates under each topic to summarize the high-scoring dialogue texts.

[0065] In one or more embodiments, the method for deleting meaningless sentences includes:

[0066] A. Segment all dialogue texts into words and use the bag-of-words model to convert them into vectors based on the frequency of each word in each sentence;

[0067] B. Mark each sentence as meaningful or meaningless;

[0068] C. Based on the vectors and annotation results of the dialogue text, a binary classification model is constructed using the logistic regression algorithm to classify all sentences in the dialogue text, thereby excluding sentences labeled as meaningless.

[0069] In simple terms, sentences in the dialogue text indicating "no intention to renew insurance" are manually annotated to indicate whether they are meaningful. Then, using a bag-of-words model combined with logistic regression, meaningful sentences are selected from each "no intention to renew insurance" dialogue text, while meaningless sentences, such as "okay" and "thank you," are excluded.

[0070] First, all dialogue texts expressing "no intention to renew insurance" are segmented into words. Then, a bag-of-words model is used to convert these words into vectors based on their frequency of occurrence in each sentence. For example, as shown in Table 1:

[0071] Table 1. Frequency of each word in the sentence text

[0072] insurance policy Renewal premium Telephone loss pity Sum Insured 2 1 1 0 0 0 1

[0073] Then, using the vectors and annotation results of the dialogue text sentences, a binary classification model is constructed using the logistic regression algorithm to classify all sentences in the dialogue text that indicate "no intention to renew insurance" and exclude meaningless sentences.

[0074] On the one hand, after deleting meaningless sentences, keywords are extracted from each sentence to analyze the differences in keyword usage on the same topic in the dialogue texts of agents with different success rates.

[0075] Preferably, after excluding meaningless sentences, the TextRank algorithm is used to extract keywords from the content spoken by the agents in each topic's dialogue text. The TextRank algorithm is a keyword extraction algorithm that can extract keywords from a single document without relying on a corpus. For example, keywords can be extracted from the retained sentences within each dialogue and each node.

[0076] On the other hand, after removing meaningless sentences, the dialogue texts of each agent with different success rates under each topic are scored to summarize the high-scoring dialogue texts.

[0077] Preferably, the steps specifically include:

[0078] A. Segment each sentence of the training text that has been pre-labeled with different score tags and use the BERT model to generate third word vectors;

[0079] B. Input the third word vector into the TextCNN model as a group for each paragraph to obtain the score for each paragraph;

[0080] C. Train the TextCNN model with training texts pre-labeled with different score tags, so that the trained TextCNN model can score all dialogue texts.

[0081] In simple terms, the content spoken by the agent in key topics within the dialogue text is scored (e.g., 1-5 points). Based on the BERT+TextCNN method, a text scoring model is built to identify high-scoring dialogue content within key topics. The specific steps are as follows:

[0082] a) Segment each sentence of the annotated dialogue text and generate word vectors using the BERT model.

[0083] b) Input the word vectors into the TextCNN model, grouped by paragraph, to obtain a score for each paragraph (e.g., 1-5 points).

[0084] c) Train the model using the scored text data, and use the trained model to score all the texts to be scored.

[0085] It should be noted that in the method described: whether the policy is renewed is determined based on whether the premium is successfully paid in the current period; the success rate is determined based on the number of text data of each agent and the renewal result; the dialogue texts of each agent with different success rates are analyzed, and at least one agent's dialogue text above the average success rate and one agent's dialogue text below the average success rate are selected.

[0086] In short, the process begins by determining whether a policy is renewed based on whether premiums were successfully paid in the current period, i.e., the renewal result. Then, the renewal success rate is determined by the number of conversations with agents who did not intend to renew and the renewal results. Therefore, the conversations of agents with high renewal success rates are highly valuable for reference, and their content can be extracted and summarized to improve overall sales script quality. For example, keywords can be extracted, and highly-scoring conversation texts for each topic can be compiled as excellent scripts for that topic. Furthermore, the conversations of agents with low renewal success rates are also valuable for reference. By comparing them, we can identify which scripts are inadequate and remind agents to avoid them during training.

[0087] In this application, high-scoring segments are summarized to extract and highlight the key points of high-scoring scripts, continuously improving and expanding the script library. Simultaneously, agents with low success rates are trained to identify shortcomings in their scripts on various topics, thereby improving renewal success rates.

[0088] For example, by summarizing high-scoring segments, we can see that excellent agents typically say in the "reason for calling" segment, "Did your policy payment fail? Let me check to see if the premium has been deposited into this card." Agents with lower success rates, however, often only say, "You have a critical illness insurance policy, and we see it hasn't been renewed yet." Therefore, through comparison, we add successful cases to our script library to improve the scripts of agents with lower success rates and increase the success rate of renewal and retention.

[0089] In one or more achievable embodiments, successful sales scripts for premium renewal are continuously summarized and accumulated based on the implementation of this invention. Comparison reveals that agents with higher success rates generally use more professional and easily understandable language to retain customers, and also offer customers multiple options to choose from. Due to the implementation of this invention, the overall success rate of premium renewal in the first half of 2021 increased by 6% compared to the same period in 2020. Therefore, this application can more easily identify excellent sales scripts and phrases in premium renewal retention, thereby improving the success rate of telephone renewals by agents.

[0090] like Figure 4The diagram shown illustrates a module schematic of a premium collection script extraction and analysis device according to an embodiment of this application. As shown, the device 400 includes:

[0091] The extraction module 401 is used to acquire the text data of the premium collection call recording and classify the text data to filter out the dialogue text of the category of no intention to renew the insurance.

[0092] The classification module 402 is used to classify each sentence in the dialogue text into topics, so as to analyze the topic missing rate and order difference between the dialogue texts of different agents with different success rates.

[0093] Analysis module 403 is used to delete meaningless sentences, extract keywords from each sentence to analyze the differences in keyword usage in the same topic among the dialogue texts of each agent with different success rates, and / or score the dialogue texts of each agent with different success rates under each topic to summarize high-scoring dialogue texts.

[0094] It should be noted that the information interaction and execution process between the modules / units of the above-mentioned device are based on the same concept as the method embodiments described in this application, and the resulting technical effects are the same as those in the method embodiments of this application. For details, please refer to the description in the method embodiments shown above in this application, and it will not be repeated here.

[0095] It should also be noted that the division of the various modules in the above device is merely a logical functional division. In actual implementation, they can be fully or partially integrated into a single physical entity, or they can be physically separated. Furthermore, these units can be implemented entirely in software via processing element calls; they can be fully implemented in hardware; or some modules can be implemented by processing element calls to software, while others are implemented in hardware. For example, each module can be a separate processing element, or it can be integrated into a chip within the above device. Alternatively, it can be stored as program code in the system's memory, and its functions can be called and executed by a processing element of the device. The implementation of other modules is similar. Moreover, these modules can be fully or partially integrated together, or they can be implemented independently. The processing element mentioned here can be an integrated circuit with signal processing capabilities. In the implementation process, each step of the above method or each of the above modules can be completed through integrated logic circuits in the hardware of the processor element or through software instructions.

[0096] For example, these modules can be one or more integrated circuits configured to implement the above methods, such as one or more Application Specific Integrated Circuits (ASICs), one or more digital signal processors (DSPs), or one or more Field Programmable Gate Arrays (FPGAs). As another example, when a module is implemented using processing element scheduler code, the processing element can be a general-purpose processor, such as a Central Processing Unit (CPU) or other processor capable of calling program code. Furthermore, these modules can be integrated together to form a System-on-a-Chip (SoC).

[0097] like Figure 5 The figure shows a schematic diagram of the structure of a computer device according to an embodiment of this application. As shown, the computer device 500 includes: a memory 501 and a processor 502; the memory 501 is used to store computer instructions; the processor 502 executes the computer instructions to implement... Figure 1 The method described.

[0098] In some embodiments, the number of memories 501 in the computer device 500 may be one or more, and the number of processors 502 may be one or more. Figure 5 Each example is taken as an instance.

[0099] In one embodiment of this application, the processor 502 in the computer device 500 will perform as follows: Figure 1 The steps described involve loading one or more instructions corresponding to the process of an application into memory 501, and then having processor 502 run the application stored in memory 501, thereby achieving the following: Figure 1 The method described.

[0100] The memory 501 may include random access memory (RAM) or non-volatile memory, such as at least one disk storage device. The memory 501 stores an operating system and operating instructions, executable modules or data structures, or subsets thereof, or extended sets thereof. The operating instructions may include various operation instructions for implementing various operations. The operating system may include various system programs for implementing various basic business processes and handling hardware-based tasks.

[0101] The processor 502 can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0102] In some specific applications, the various components of the computer device 500 are coupled together through a bus system, which may include not only a data bus but also a power bus, a control bus, and a status signal bus. However, for the sake of clarity, in... Figure 5 All kinds of buses are referred to as bus systems.

[0103] In one embodiment of this application, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the following: Figure 1 The method described.

[0104] At any possible level of technical detail, this application can be a system, method, and / or computer program product. A computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for causing a processor to implement various aspects of this application.

[0105] Computer-readable storage media can be tangible devices capable of holding and storing instructions for use by an instruction execution device. Computer-readable storage media can be, for example, (but not limited to) electrical storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of computer-readable storage media include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory sticks, floppy disks, mechanical encoding devices, such as punch cards or recessed protrusions storing instructions thereon, and any suitable combination of the foregoing. The computer-readable storage media used herein are not to be construed as transient signals themselves, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (e.g., light pulses through fiber optic cables), or electrical signals transmitted through wires.

[0106] The computer-readable program described herein can be downloaded from a computer-readable storage medium to various computing / processing devices, or downloaded via a network, such as the Internet, local area network, wide area network, and / or wireless network, to an external computer or external storage device. The network may include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers. A network adapter card or network interface in each computing / processing device receives computer-readable program instructions from the network and forwards those instructions to the computer-readable storage medium stored in the respective computing / processing device.

[0107] The computer program instructions used to perform the operations of this application may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, status setting data, integrated circuit configuration data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Smalltalk, C++, etc., and procedural programming languages ​​such as the "C" language or similar programming languages. The computer-readable program instructions may be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or may be connected to an external computer (e.g., via the Internet using an Internet service provider). In some embodiments, electronic circuits, such as programmable logic circuits, field-programmable gate arrays (FPGAs), or programmable logic arrays (PLAs), are personalized by utilizing the status information of the computer-readable program instructions. These electronic circuits can execute the computer-readable program instructions to implement various aspects of this application.

[0108] In summary, the method, apparatus, equipment, and medium for extracting and analyzing premium collection scripts provided in this application acquire text data from recorded premium collection calls, classify the text data to filter out dialogue texts belonging to the category of no intention to renew insurance; classify each sentence in the dialogue text by topic to analyze the topic missing rate and order differences between dialogue texts of different agents with different success rates; delete meaningless sentences, extract keywords from each sentence to analyze the wording differences of keywords in the same topic in dialogue texts of different agents with different success rates, and / or score the dialogue texts of different agents with different success rates under each topic to summarize high-scoring dialogue texts.

[0109] This application effectively overcomes the various shortcomings of the prior art and has high industrial application value.

[0110] The above embodiments are merely illustrative of the principles and effects of this application and are not intended to limit the invention. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of this application. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in this invention should still be covered by the claims of this application.

Claims

1. A method for extracting and analyzing premium collection scripts, characterized in that, The method includes: Obtain the text data of the premium collection call recording, and classify the text data to filter out the dialogue text of the category of no intention to renew the policy; The method for classifying the text data to filter out dialogue texts in the category of no intention to renew insurance includes: The training text, which is pre-labeled with different classification tags, is segmented into words and the first word vector is generated using the BERT model. The first word vector is used as input to the TextCNN model to train the TextCNN model; The text data is initially screened and classified using a trained TextCNN model to select dialogue texts with the classification label of "no intention to renew insurance". Each sentence in the dialogue text is classified by topic to analyze the topic missing rate and order differences among the dialogue texts of different agents with different success rates. The method for classifying each sentence in the dialogue text by topic includes: Each sentence of the training text, which is pre-labeled with different topic tags, is segmented into words, and a second word vector is generated using the BERT model; The second word vector is input into the sentence-level Bi-LSTM model as a group for each sentence to obtain the sentence vector for each sentence; Each sentence vector and the interlocutor's identity information are input into the Bi-LSTM model in the order of the dialogue to obtain the conversation vector corresponding to each sentence; The session vectors are converted into the probability that each sentence is the starting sentence of a topic paragraph using a fully connected SoftMax layer; The probability is used as input to the CRF layer and the probability value is optimized to obtain the corrected final probability; The Bi-LSTM model is trained using training texts pre-labeled with different topic tags, so that the trained Bi-LSTM model can be used to classify each sentence in the dialogue text into a topic. Delete meaningless sentences, extract keywords from each sentence to analyze the differences in keyword usage in the same topic among the dialogue texts of each agent with different success rates, and / or score the dialogue texts of each agent with different success rates under each topic to summarize the high-scoring dialogue texts. The method for scoring the dialogue texts of each agent with different success rates under each topic to summarize the high-scoring dialogue texts includes: Each sentence of the training text, which is pre-labeled with different score tags, is segmented into words, and a third word vector is generated using the BERT model; The third word vector is input into the TextCNN model as a group for each paragraph to obtain a score for each paragraph; The TextCNN model is trained using training texts pre-labeled with different score tags, so that the trained TextCNN model can score all dialogue texts.

2. The method according to claim 1, characterized in that, The topic categories include any one or more of the following: opening remarks, stating the reason for the call, inquiring about the reasons for not renewing the policy, stating the advantages and disadvantages, and closing remarks.

3. The method according to claim 1, characterized in that, The methods for deleting meaningless sentences include: All dialogue texts are segmented into words, and the bag-of-words model is used to convert them into vectors according to the frequency of each word in each sentence; Each sentence is labeled as meaningful or meaningless; Based on the vectors and annotation results of the dialogue text, a binary classification model is constructed using the logistic regression algorithm to classify all sentences in the dialogue text, thereby excluding sentences labeled as meaningless.

4. The method according to claim 1, characterized in that, The method includes: Whether the policy is renewed is determined based on whether the premium was successfully paid in the current period; The success rate is determined based on the amount of text data from each agent and the renewal results. Analyze the dialogue texts of each agent with different success rates, and select at least one agent's dialogue text with a success rate above the average and one agent's dialogue text with a success rate below the average.

5. A device for extracting and analyzing insurance premium collection scripts, characterized in that, The device includes: An extraction module is used to acquire text data from recorded premium collection calls and classify the text data to filter out dialogue texts in the category of "no intention to renew insurance". The method for classifying the text data to filter out dialogue texts in the category of "no intention to renew insurance" includes: segmenting training texts pre-labeled with different classification tags and generating first word vectors using a BERT model; using the first word vectors as input to a TextCNN model to train the TextCNN model; and using the trained TextCNN model to perform initial classification of the text data to select dialogue texts with the classification tag of "no intention to renew insurance". A classification module is used to classify each sentence in the dialogue text into topics, so as to analyze the topic missing rate and order differences between dialogue texts of different agents with different success rates. The method for classifying each sentence in the dialogue text into topics includes: segmenting each sentence of the training text pre-labeled with different topic tags and generating a second word vector using a BERT model; inputting the second word vectors into a sentence-level Bi-LSTM model as a group of sentences to obtain a sentence vector for each sentence; inputting each sentence vector and the interlocutor's identity information into the Bi-LSTM model in the dialogue order to obtain a conversation vector corresponding to each sentence; using a fully connected SoftMax layer to convert the conversation vector into the probability that each sentence is the starting sentence of a topic paragraph; using the probability as the input of a CRF layer and optimizing the probability value to obtain the corrected final probability; training the Bi-LSTM model with the training text pre-labeled with different topic tags, so that the trained Bi-LSTM model can be used to classify each sentence in the dialogue text into topics. The analysis module is used to delete meaningless sentences, extract keywords from each sentence to analyze the differences in keyword usage among agents with different success rates in the same topic, and / or score the dialogue texts of agents with different success rates under each topic to summarize high-scoring dialogue texts; the method of scoring the dialogue texts of agents with different success rates under each topic to summarize high-scoring dialogue texts includes: segmenting each sentence of the training text pre-labeled with different score labels and generating a third word vector using the BERT model; inputting the third word vectors into the TextCNN model as a group for each paragraph to obtain the score for each paragraph; training the TextCNN model with the training text pre-labeled with different score labels so that the trained TextCNN model can score all dialogue texts.

6. A computer device, characterized in that, The device includes: a memory and a processor; the memory is used to store computer instructions; the processor executes the computer instructions to implement the method as described in any one of claims 1 to 4.

7. A computer-readable storage medium, characterized in that, The device stores computer instructions that, when executed, perform the method as described in any one of claims 1 to 4.

Images