Call auxiliary database updating method and device, equipment, storage medium and program product

Abstract

The invention provides a call auxiliary database updating method and device, equipment, a storage medium and a program product, and relates to the technical field of data processing. The method comprises the following steps: in response to determining that the text length of a historical call text is greater than a preset text length, inputting the historical call text into a large language model for simplification processing to obtain a target call text; inputting the target call text, the historical intention table and a historical customer expression table corresponding to the historical intention table into a large language model for processing to obtain a target customer expression and a target customer intention corresponding to the target customer expression; and updating the historical customer expression table based on the target customer expression to obtain a target customer expression table, and updating the historical intention table based on the target customer intention to obtain a target intention table. According to the method, the target customer expression and the target customer intention can be efficiently obtained, and the updating efficiency of the call auxiliary database is effectively improved.

Description

Technical Field

[0001] The present application relates to the field of data processing technology, and in particular to a call assistance database updating method, device, equipment, storage medium and program product. Background Art

[0002] The existing call assistance system can analyze the conversation content between the agent and the customer in real time, automatically identify the customer's actual intention, and match the corresponding excellent words for the agent's reference. The realization of this function mainly relies on the call assistance database, which includes a customer intention list and a customer intention similar question list. As the business environment continues to change, the customer's intention and expression methods are also constantly updated. Therefore, the two lists in the customer intention similar question database also need to be updated frequently to ensure that the recommendation function always maintains efficiency and accuracy.

[0003] At present, the update of call auxiliary data mainly relies on manual work, that is, excellent agents or experts read and analyze massive historical call records, extract new customer intentions and expressions from them for updating. However, this update method is inefficient and cannot meet rapidly changing business needs and customer expectations. Summary of the invention

[0004] The present application provides a call assistance database update method, device, equipment, storage medium and program product to solve the problem of low efficiency of existing call assistance database update methods.

[0005] In a first aspect, the present application provides a method for updating a call assistance database, comprising:

[0006] Get historical call texts;

[0007] In response to determining that the text length of the historical call text is greater than a preset text length, inputting the historical call text into a large language model for simplification processing to obtain a target call text;

[0008] Inputting the target call text, the historical intent table, and the historical customer expression table corresponding to the historical intent table into a large language model for processing to obtain a target customer expression and a target customer intention corresponding to the target customer expression;

[0009] The historical customer expression table is updated based on the target customer expression to obtain a target customer expression table, and the historical intent table is updated based on the target customer intent to obtain a target intent table.

[0010] In a possible design, the inputting the historical call text into a large language model for simplification processing to obtain a target call text includes:

[0011] Splitting the historical call text based on the large language model and preset prompt words to obtain multiple initial paragraphs;

[0012] Based on the large language model, the contents of multiple initial paragraphs are refined and summarized to generate multiple summaries;

[0013] The contents of the multiple initial paragraphs are merged with the multiple summaries based on a preset merging rule to obtain a target call text.

[0014] In a possible design, the merging of the contents of the multiple initial paragraphs with the multiple summaries based on a preset merging rule to obtain a target call text includes:

[0015] Merge the content of the current initial paragraph with the summaries corresponding to other initial paragraphs to obtain a target paragraph;

[0016] Merge multiple target paragraphs to obtain the target call text.

[0017] In a possible design, the target call text, the historical intent table, and the historical customer expression table corresponding to the historical intent table are input into a large language model for processing to obtain a target customer expression and a target customer intent corresponding to the target customer expression, including:

[0018] Performing semantic extraction on the target call text, the historical intent table, and the historical customer expression table based on a large language model to determine the target customer expression;

[0019] Matching the target customer expression with the historical intent table based on a large language model to determine whether there is a historical customer intent that matches the target customer expression;

[0020] In response to determining that there is a historical customer intent that matches the target customer expression, the historical customer intent is determined as the target customer intent.

[0021] In one possible design, it also includes:

[0022] In response to determining that no historical customer intent matches the target customer expression, a target customer intent is generated based on a large language model and the target customer expression.

[0023] In a possible design, the target call text includes multiple paragraphs, and the multiple paragraphs are obtained by simplifying the historical call text with a large language model;

[0024] Before the semantic extraction of the target call text, the historical intent table and the historical customer expression table based on the large language model to determine the target customer expression, the method further includes:

[0025] splicing the contents of multiple paragraphs with the historical intent table and the historical customer expression table one by one to obtain the text to be input;

[0026] The text to be input is input into the large language model.

[0027] In a possible design, the updating of the historical customer expression table based on the target customer expression to obtain a target customer expression table, and the updating of the historical intent table based on the target customer intent to obtain a target intent table, include:

[0028] Displaying the target customer statement and the target customer intention corresponding to the target customer statement;

[0029] In response to receiving a confirmation operation of the user on the target customer statement through an input device, updating the target customer statement to the historical customer statement table to obtain a target customer statement table;

[0030] In response to receiving a determination operation of the target customer intention by the user through an input device, the target customer intention is updated to the historical intention table to obtain a target intention table.

[0031] In one possible design, it also includes:

[0032] Sorting the target customer statements according to a preset sorting rule;

[0033] The sorted target customer expressions and the target customer intentions corresponding to the target customer expressions are displayed.

[0034] In a possible design, the target customer statements are sorted according to a preset sorting rule, including:

[0035] The target customer expressions are sorted according to semantic similarity, repetition frequency and meaning. Figure 1 Any one or more combinations of the degree of consistency are used to sort.

[0036] In one possible design, it also includes:

[0037] In response to determining that the text length of the historical call text is less than or equal to a preset text length, the historical call text, the historical intent table, and the historical customer statement table corresponding to the historical intent table are input into a large language model for semantic extraction and intent matching to obtain a target customer statement and a target customer intent corresponding to the target customer statement.

[0038] In a second aspect, the present application provides a call assistance database updating device, comprising:

[0039] The acquisition module is used to obtain historical call texts;

[0040] a simplification module, configured to, in response to determining that the text length of the historical call text is greater than a preset text length, input the historical call text into a large language model for simplification processing to obtain a target call text;

[0041] A processing module, configured to input the target call text, the historical intent table, and the historical customer expression table corresponding to the historical intent table into a large language model for processing, so as to obtain a target customer expression and a target customer intent corresponding to the target customer expression;

[0042] An updating module is used to update the historical customer expression table based on the target customer expression to obtain a target customer expression table, and to update the historical intent table based on the target customer intent to obtain a target intent table.

[0043] In a third aspect, the present application provides an electronic device, comprising: a processor, and a memory communicatively connected to the processor;

[0044] The memory stores computer-executable instructions;

[0045] The processor executes the computer-executable instructions stored in the memory to implement the method as described in any one of the first aspects.

[0046] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium, wherein the computer-readable storage medium stores computer-executable instructions, and when the computer-executable instructions are executed by a processor, they are used to implement the method as described in any one of the first aspects.

[0047] In a fifth aspect, an embodiment of the present application provides a computer program product, including a computer program, which, when executed by a processor, implements the method described in any one of the first aspects.

[0048] The call auxiliary database update method, device, equipment, storage medium and program product provided by the present application obtain historical call texts; in response to determining that the text length of the historical call texts is greater than the preset text length, the historical call texts are input into the large language model for simplified processing to obtain the target call text, thereby reducing the text complexity and improving the efficiency and accuracy of the subsequent large language model calculations; the target call texts, the historical intent table and the historical customer expression table corresponding to the historical intent table are input into the large language model for processing to obtain the target customer expression and the target customer intention corresponding to the target customer expression; the historical customer expression table is updated based on the target customer expression to obtain the target customer expression table, and the historical intention table is updated based on the target customer intention to obtain the target intention table. Compared with the prior art, the method of the present application can obtain the target customer expression and the target customer intention more efficiently, effectively improving the update efficiency of the call auxiliary database. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present application and, together with the description, serve to explain the principles of the present application.

[0050] Figure 1 An application scenario diagram corresponding to the call assistance database update method provided in an embodiment of the present application;

[0051] Figure 2 A flowchart of a method for updating a call assistance database provided in an embodiment of the present application;

[0052] Figure 3 A flowchart of a method for updating a call assistance database provided in another embodiment of the present application;

[0053] Figure 4 A schematic diagram of the structure of a call assistance database updating device provided in one embodiment of the present application;

[0054] Figure 5 This is a diagram showing an example structure of an electronic device provided in one embodiment of the present application.

[0055] The above drawings have shown clear embodiments of the present application, which will be described in more detail later. These drawings and text descriptions are not intended to limit the scope of the present application in any way, but to illustrate the concept of the present application to those skilled in the art by referring to specific embodiments. DETAILED DESCRIPTION

[0056] Exemplary embodiments will be described in detail herein, examples of which are shown in the accompanying drawings. When the following description refers to the drawings, the same numbers in different drawings represent the same or similar elements unless otherwise indicated. The implementations described in the following exemplary embodiments do not represent all implementations consistent with the present application. Instead, they are merely examples of devices and methods consistent with some aspects of the present application as detailed in the appended claims.

[0057] In order to clearly understand the technical solution of the present application, the solution of the prior art is first introduced in detail.

[0058] In the financial field, excellent agents can accurately grasp the potential needs and true intentions of customers during calls with their rich experience, keen insight and excellent communication skills, thereby providing customized financial product recommendations and effectively enhancing customers' sense of identity and trust. However, the agent team is often large in size and has high staff turnover. Therefore, in order to ensure the professionalism of the agents, they need to be trained frequently.

[0059] In order to reduce the cost of frequent training of agents, currently, a call assistance system is introduced to assist agents in providing professional services to customers. Specifically, when an agent is talking to a customer, the call assistance system can obtain and analyze the conversation content between the agent and the customer in real time, identify the customer's intention, and then provide the agent with a response that matches the customer's intention. Among them, the call assistance system mainly relies on the call assistance database for intent matching. Specifically, the call assistance database includes a customer intention list and a customer intention similar question list.

[0060] As the business environment continues to change, customers' intentions and expressions are constantly updated. Therefore, the two lists in the customer intention similarity database also need to be updated frequently to ensure that the recommendation function always maintains efficiency and accuracy. At present, the update of the call assistance database mainly relies on manual completion, that is, excellent agents or experts read and analyze massive historical call records, and combine their own business knowledge and experience to extract new customer intentions and expressions from historical call records for updating. Although the above method can effectively and accurately update the call assistance database, this update method is time-consuming and inefficient, and it is difficult to meet rapidly changing business needs and customer expectations.

[0061] Therefore, in the face of the technical problems of the above-mentioned prior art, in order to improve the updating efficiency of the call assistance database. First, the long text of the massive call text is simplified through the machine learning model to reduce the difficulty of subsequent model processing; then the simplified text is input into the machine learning model to extract the customer expression, and the extracted customer expression is matched with the existing customer intent table to finally determine the customer expression and the corresponding customer intent, and then update the customer expression and customer intent to the call assistance database. Therefore, by automatically extracting customer expressions in call texts and matching customer intent, the updating efficiency of the call assistance database can be effectively improved.

[0062] Figure 1 An application scenario diagram corresponding to the call assistance database update method provided in an embodiment of the present application, such as Figure 1 As shown in , the application scenario provided by this embodiment includes: a user terminal 10, a database update device 11 and a database 12. Specifically, when the call auxiliary database needs to be updated, the user terminal 10 sends an update instruction to the database update device 11; after receiving the update instruction, the database update device 11 obtains the historical call text, the historical intention table and the historical customer expression table from the database 12; the database update device 11 simplifies the historical call text based on the large language model to obtain the target call text; the target call text, the historical intention table and the historical customer expression table are input into the large language model for processing to obtain the target customer expression and the target customer intention; the historical intention table and the historical customer expression table are updated based on the target customer expression and the target customer intention to obtain the target customer expression table and the target intention table. Finally, the database update device 11 sends the target customer expression table and the target intention table to the database 12, thereby completing the update of the database 12.

[0063] The technical solution of the present application and how the technical solution of the present application solves the above-mentioned technical problems are described in detail below with specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments. The embodiments of the present application will be described below in conjunction with the accompanying drawings.

[0064] In the several embodiments provided in the present application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic, for example, the division of modules is only a logical function division, and there may be other division methods in actual implementation, such as multiple modules can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interfaces, devices or modules, which can be electrical, mechanical or other forms.

[0065] Figure 2 A flowchart of a method for updating a call assistance database provided in an embodiment of the present application is shown in FIG. Figure 2 As shown, the execution subject of this embodiment is a call assistance database update device, which can be implemented by a computer program or a medium storing relevant computer programs, such as a USB flash drive and / or a CD, etc.; or, it can also be implemented by a physical device integrated or installed with relevant computer programs, such as a chip, an electronic device, etc. The electronic device can be a computer or a server, etc. The call assistance database update method provided in this embodiment includes the following steps:

[0066] Step S201, obtaining historical call texts.

[0067] Specifically, when updating the call assistance database, the historical call texts stored in the call assistance database are first obtained for subsequent analysis and processing.

[0068] It should be noted that the historical call text is a text recording the historical conversation content between the agent and the customer. Specifically, the historical call text is obtained from the historical call records through a speech-to-text conversion tool.

[0069] Optionally, the historical call text may be a call record of the past month, or a call record of the past week. Optionally, a time range for acquiring the text may be set according to demand or processing capability to acquire historical call texts of a specific time range.

[0070] Step S202: In response to determining that the text length of the historical call text is greater than the preset text length, the historical call text is input into the large language model for simplification processing to obtain the target call text.

[0071] Specifically, after obtaining the historical call text, it is determined whether the text length of the historical call text is greater than the preset text length. When the text length of the historical call text exceeds the preset text length, in order to reduce the difficulty of subsequent processing, the historical call text is input into the large language model for simplification processing, thereby obtaining the simplified target call text.

[0072] The preset text length is a text length preset manually and can be adjusted according to actual needs. Optionally, the preset text length can be a preset number of text characters or a preset number of text paragraphs, etc., which is not limited in this embodiment.

[0073] Among them, the big language model is a machine learning model that can understand and generate natural language text. It can analyze and reconstruct the input text to generate a more concise and easy-to-understand version.

[0074] At present, large language models have a problem of reduced ability for complex reasoning tasks with long texts and multiple facts. Therefore, in this embodiment, the call record text is simplified while maintaining the necessary information, thereby reducing the difficulty of subsequent large language model processing and improving the mining ability of the large language model.

[0075] Optionally, the large language model has a variety of options for the specific way of simplifying the historical call text. For example, keyword extraction: by analyzing the historical call text, extract key information or keywords, retain only these keywords and their related important information, and remove redundant and minor content; sentence compression: compress long sentences, remove unnecessary modifications and redundant information, retain the core meaning of the sentence, and make the expression more concise and clear; context fusion: analyze the context information in the text, merge related content together, avoid repeating the same information, and make the text more compact; text splitting, guide the large model to split the historical call text into multiple different paragraphs through preset prompt words. Specifically, in this embodiment, the method of simplifying the historical call text is not specifically limited.

[0076] The target call text is the call text simplified by the large language model.

[0077] Step S203: input the target call text, the historical intent table, and the historical customer expression table corresponding to the historical intent table into the large language model for processing to obtain the target customer expression and the target customer intention corresponding to the target customer expression.

[0078] Specifically, after obtaining the target call text, the target call text, the historical intent table and the historical customer expression table are input into the large language model for processing, thereby obtaining the target customer expression and the target customer intent.

[0079] Among them, the historical intent table is a table that stores historical customer intents. The table records various customer intents identified in historical customer conversations and their corresponding labels. Specifically, customer intents may include inquiring for information, making complaints, making suggestions, requesting help, etc.

[0080] The historical customer expression table is a table storing historical customer expressions, which records various expressions used by customers in historical customer conversations and the intentions corresponding to these expressions.

[0081] It should be noted that the historical intent table and the historical customer expression table provide rich background information and learning materials for the large language model, alleviating the problem that the large language model cannot well understand the common problem of inexpressiveness in call texts (such as excuses, refusals, unclear expressions, etc.), thereby enhancing the understanding and prediction capabilities of the large language model when processing target call texts. Specifically, the role of the historical intent table and the historical customer expression table is mainly reflected in the following aspects: context understanding, the historical intent table and the historical customer expression table provide background information of past customer interactions. Through these historical data, the large language model can better understand the context in the target call text, thereby accurately identifying customer intent; pattern recognition, historical data helps the large language model identify common customer expression patterns and corresponding intents. This pattern recognition capability can help the large language model more accurately match the target customer expression with its intent; semantic association, by analyzing historical customer expressions and corresponding customer intents, the large language model can learn the semantic association between different expressions, which helps to process diverse customer expressions and improve the ability to understand the target customer expressions.

[0082] Optionally, the target call text, the historical intent table, and the historical customer expression table are integrated into one or more input formats so that the large language model can process them. Optionally, the historical intent table and the historical customer expression table are added to the front or back of the target call text, and are presented in the form of a list or key-value pair. Optionally, a JSON object is created, which includes fields such as call text, historical intent, and historical customer expression, and the value of each field can be a string, a list, or other JSON object.

[0083] The target customer description is one or more customer descriptions similar to the target call content generated by the large language model based on the input target call text and the historical customer description table.

[0084] The target customer intent is the customer intent matched or generated based on the target call text, the historical intent table, and the historical customer expression table. The customer intent is usually one or more tags, indicating the goal that the customer wants to achieve or the emotion that the customer expresses during the call.

[0085] Specifically, the processing content of the large language model for the input target call text, historical intent table and historical customer expression table includes but is not limited to:

[0086] Text understanding and analysis: Parse the input target call text to identify the vocabulary, phrases, sentence structure, etc. to understand the basic meaning of the text; and combine the historical intent table and the historical customer expression table to conduct contextual analysis of the target call text to understand the customer's true intentions and needs in the current call.

[0087] Intent recognition and matching: Identify the customer's intent based on the input call text content; and match the identified intent with the historical intent table to find the intent that best meets the customer's current needs. If the match is successful, the corresponding intent can be directly output; if the match is unsuccessful, further analysis and processing may be required.

[0088] Determine target customer expressions and target customer intent: Generate target customer expressions that meet customer needs based on the identified intent and context information, and determine the target customer intent corresponding to the target customer expressions.

[0089] Step S204: updating the historical customer expression table based on the target customer expression to obtain a target customer expression table, and updating the historical intent table based on the target customer intent to obtain a target intent table.

[0090] Specifically, after obtaining the target customer expression and the target customer intention, the historical customer expression table and the historical intention table are updated to obtain the target customer expression table and the target intention table.

[0091] Optionally, when updating the historical customer expression table and the historical intent table, the target customer expression may be directly added to the historical customer expression table, and the target customer intent may be directly added to the historical intent table.

[0092] Optionally, when updating the historical customer expression table and the historical intent table, the target customer expression can be compared with the historical customer expression table to find similar or related historical customer expressions. If there is a difference between the target customer expression and the historical customer expression, the historical customer expression table is updated according to the target customer expression. If the target customer expression is new, it is added to the historical customer expression table and a new classification or label is created. Optionally, the identified target customer intent can be compared with the historical intent table to find historical intents that are similar or related to the target customer intent. Based on the target customer intent and the historical intent, it is determined whether to add a new intent, modify the weight or classification of the old intent, etc.

[0093] Optionally, historical customer expressions that are similar or related to the target customer expression table are found manually or by an algorithm, and historical intentions that are similar or related to the target customer intentions are determined manually or by an algorithm.

[0094] Optionally, establish a connection between target customer statements and target customer intentions so that the customer's real needs can be quickly located and understood in subsequent analysis and decision-making.

[0095] The method for updating the call auxiliary database provided in this embodiment obtains the historical call text; in response to determining that the text length of the historical call text is greater than the preset text length, the historical call text is input into the large language model for simplification processing to obtain the target call text; the target call text, the historical intention table and the historical customer expression table corresponding to the historical intention table are input into the large language model for processing to obtain the target customer expression and the target customer intention corresponding to the target customer expression; the historical customer expression table is updated based on the target customer expression to obtain the target customer expression table, and the historical intention table is updated based on the target customer intention to obtain the target intention table. The method provided in this embodiment reduces the text complexity and improves the efficiency and accuracy of the subsequent large language model calculation by simplifying the historical call text, using the large language model to process the input target call text, the historical intention table and the historical customer expression table, and updates the historical customer expression table and the historical customer intention based on the obtained target customer expression and target customer intention. Compared with the prior art, the method of this embodiment can obtain the target customer expression and the target customer intention more efficiently, and effectively improves the update efficiency of the call auxiliary database.

[0096] For example, in order to better understand the call assistance database update method provided in this embodiment, a specific example is provided below:

[0097] First, get the historical call texts.

[0098] "Hello, I am a customer service representative from XX Finance. I understand that you were interested in our financial products before. I would like to introduce our loan products to you in detail. It has a very low interest rate of only 3.5%, and a flexible repayment period of up to 30 years. Do you have any questions about our products or need further information?"

[0099] Secondly, determine whether the text length of the historical call text is greater than the preset text length. Assuming that the preset text length is 300 words, the above historical call text needs to be input into the large language model for simplification.

[0100] Specifically, the simplified target call text is: "The customer is interested in financial products. Introduce loan products: interest rate 3.5%, repayment period up to 30 years, ask about customer needs."

[0101] Table 1: Example table of historical intent

[0102]

[0103] Table 2: Example of historical customer statements

[0104]

[0105] Next, the target call text, the historical intent example table, and the historical customer expression example table corresponding to the historical intent example table are input into the large language model for processing. The historical intent example table is shown in Table 1, and the historical customer expression example table is shown in Table 2.

[0106] Specifically, after processing by the large language model, the target customer’s expression is: “Interested in financial products, asking about loan terms and repayment period”; the target customer’s intention is: to inquire about financial product information.

[0107] Finally, based on the target customer's statement, the historical customer statement example table is updated to obtain the target customer statement example table; based on the target customer's intention, the historical intention example table is updated to obtain the target intention example table. Since the target customer's intention obtained in this example is an existing customer's intention, the target intention example table and the historical intention example table are the same and are no longer displayed. Specifically, the target customer statement example table is shown in Table 3.

[0108] Table 3: Example of target customer description

[0109]

[0110] Thus, an update of the historical customer expression example table and the historical intent example table is completed. It should be noted that the above example is only a specific case to assist in understanding the method of this embodiment, and it does not represent the only case or all contents of the method in actual application.

[0111] As an optional implementation, based on any of the above embodiments, the historical call text is input into the large language model for simplified processing to obtain the target call text, including: splitting the historical call text based on the large language model and preset prompt words to obtain multiple initial paragraphs; refining and summarizing the contents of the multiple initial paragraphs based on the large language model to generate multiple summaries; merging the contents of the multiple initial paragraphs with the multiple summaries based on preset merging rules to obtain the target call text.

[0112] Specifically, when the historical call text is input into the large language model for simplification processing, the historical call text is first split based on the large language model and preset prompt words to obtain multiple initial paragraphs, and then the multiple initial paragraphs are refined and summarized based on the large language model to obtain multiple summaries corresponding to the contents of the initial paragraphs; finally, the contents of the multiple initial paragraphs are merged with the multiple summaries based on the preset merging rules to obtain the target call text.

[0113] Among them, the preset prompt words are pre-set keywords, phrases or sentence structures, and their role is to guide the large language model to understand the historical call text and divide it into different multiple initial paragraphs. The preset prompt words may include paragraph separators, topic transition signs, etc., which help the large language model to more accurately identify the structure of the historical call text. For example, the preset prompt words can be specific stages or topics closely related to the content of the call, such as opening remarks, self-introduction, introduction of products, introduction of product features, answering customer questions, and transaction facilitation; or, the preset prompt words can be general prompt words such as "question", "answer", and "viewpoint". In this embodiment, the preset prompt words are not specifically limited. In actual applications, the preset prompt words can be flexibly selected and adjusted according to the specific call content and goals to ensure that the model can accurately understand and divide the historical call text.

[0114] Optionally, the historical call text is split based on the large language model and preset prompt words, so that the historical call text is decomposed into multiple initial paragraphs, each of which contains relatively independent information or topics. Specifically, there are many ways to split the historical call text, including but not limited to: splitting based on keywords, the large language model identifies keywords that match the preset prompt words and splits the historical call text into different paragraphs based on this; splitting based on sentence structure, the large language model analyzes the grammatical structure of the sentence, such as subject, predicate, object, etc., and divides the paragraphs according to the changes in the structure; splitting based on semantics, the large language model understands the semantic content of the historical call text, and splits the paragraphs according to the conversion of topics or changes in logical relationships. In this embodiment, the splitting method is not specifically limited. In actual applications, the splitting method is flexibly selected and adjusted to achieve the best splitting effect.

[0115] Optionally, the contents of multiple initial paragraphs are refined and summarized based on the large language model, so that redundant information can be removed and the core content or key information of each initial paragraph can be extracted. In this way, through refinement and summarization, each initial paragraph is converted into a summary, which retains the main information of the original initial paragraph, but is more refined and easy to understand. Specifically, there are many ways to refine the content of the initial paragraph, including but not limited to: extracting key information, the large language model identifies and extracts key information in the initial paragraph, such as time, place, people, events, etc., to form a summary; using a summary template, based on a preset summary module, the large language model generates a summary that meets the requirements according to the summary template. In this embodiment, the generation method of the summary is not specifically limited. In practical applications, the summary generation method can be flexibly selected and adjusted.

[0116] The preset merging rule is a preset rule for merging the initial paragraph and the abstract.

[0117] Optionally, the contents of multiple initial paragraphs are merged with multiple summaries based on a preset merging rule to obtain a target call text, which is more concise than the historical call text and is convenient for subsequent processing and analysis. Optionally, there are multiple ways to merge the initial paragraphs with the summaries, such as: retaining the order of the original initial paragraphs, inserting the summaries into the appropriate positions in the original initial paragraphs, or completely replacing the original initial paragraphs with the summaries. In this embodiment, no specific limitation is imposed on the way of merging the initial paragraphs with the summaries.

[0118] The call assistance database update method provided in this embodiment inputs the historical call text into the large language model for simplified processing to obtain the target call text, including: splitting the historical call text based on the large language model and preset prompt words to obtain multiple initial paragraphs; refining and summarizing the contents of the multiple initial paragraphs based on the large language model to generate multiple summaries; merging the contents of the multiple initial paragraphs with the multiple summaries based on the preset merging rules to obtain the target call text. The method of this embodiment can accurately and efficiently split the historical call text and refine and summarize the core content and remove redundant information through the large language model and preset prompt words; merge the initial paragraphs with the summaries based on the preset merging rules to generate a concise and clear target call text for subsequent processing and analysis.

[0119] As an optional implementation, based on any of the above embodiments, the contents of multiple initial paragraphs are merged with multiple summaries based on preset merging rules to obtain a target call text, including: merging the content of the current initial paragraph with the summaries corresponding to other initial paragraphs to obtain a target paragraph; merging multiple target paragraphs to obtain a target call text.

[0120] Specifically, when merging the contents of multiple initial paragraphs with multiple summaries based on preset merging rules to obtain a target call text, firstly merge the current initial paragraph content with the summaries corresponding to other initial paragraphs to obtain a target paragraph having the original text of the current initial paragraph and summaries of other paragraphs; then merge the multiple target paragraphs to obtain the target call text.

[0121] It should be noted that before merging the initial paragraphs and the abstracts, the relationship between each initial paragraph and its corresponding abstract needs to be determined. Optionally, tags are used to clarify the relationship between the initial paragraphs and the corresponding abstracts, that is, specific identifiers or tags are assigned to the initial paragraphs and the abstracts; or, the semantic relevance analysis capability of a large language model is relied upon to automatically determine the relationship between the initial paragraphs and their corresponding abstracts.

[0122] The target paragraph is a paragraph including the original text of an initial paragraph and the summaries of other paragraphs. Optionally, the target paragraph may include the original text of the current initial paragraph and the summaries corresponding to the initial paragraphs adjacent to it, or the target paragraph includes the original text of the current initial paragraph and the summaries corresponding to all the remaining initial paragraphs. In this embodiment, the specific composition of the target paragraph is not limited.

[0123] Optionally, when merging the original text of the initial paragraph with the summary, a variety of strategies can be adopted, such as inserting summary content, semantic fusion of the initial paragraph and the summary, etc., with the aim of effectively integrating the key information in the summary into the current paragraph to form a more complete and refined target paragraph.

[0124] Optionally, before the content of the current initial paragraph is merged with the summaries corresponding to other initial paragraphs, the current initial paragraph is first analyzed and judged to determine whether to retain the initial paragraph. For example, taking the preset prompt words of opening remarks, self-introduction, product introduction, product feature introduction, answering customer questions, and transaction facilitation to split the historical call text into multiple initial paragraphs, the content of the initial paragraph corresponding to the opening remarks does not involve the actual recommended content, so it does not need to be retained, so the initial paragraph corresponding to the opening remarks is deleted before the initial paragraph and the summary are merged to reduce the difficulty of subsequent processing.

[0125] Optionally, after obtaining multiple target paragraphs, the multiple target paragraphs are further merged to obtain a complete target call text. Specifically, the target paragraph merging may involve adjusting the order between target paragraphs, removing duplicate information, etc., and following a certain logical structure during merging, such as time sequence, cause and effect, topic classification, etc., to ensure the coherence and orderliness of the target call text.

[0126] The call auxiliary database updating method provided in this embodiment merges the contents of multiple initial paragraphs with multiple summaries based on a preset merging rule to obtain a target call text, including: merging the contents of the current initial paragraph with the summaries corresponding to other initial paragraphs to obtain a target paragraph; merging multiple target paragraphs to obtain a target call text. The method of this embodiment can effectively ensure the integrity and refinement of the target call text by merging the current initial paragraph with the summaries of other paragraphs to form a target paragraph and then obtaining the target call text.

[0127] As an optional implementation, based on any of the above embodiments, the target call text, the historical intent table, and the historical customer statement table corresponding to the historical intent table are input into the large language model for processing to obtain the target customer statement and the target customer intention corresponding to the target customer statement, including: performing semantic extraction on the target call text, the historical intent table, and the historical customer statement table based on the large language model to determine the target customer statement; matching the target customer statement with the historical intent table based on the large language model to determine whether there is a historical customer intention that matches the target customer statement; in response to determining that there is a historical customer intention that matches the target customer statement, determining the historical customer intention as the target customer intention.

[0128] Specifically, when the target call text, the historical intent table, and the historical customer expression table corresponding to the historical intent table are input into the large language model for processing to obtain the target customer expression and the target customer intention corresponding to the target customer expression, first, semantic extraction is performed on the target call text, the historical intent table, and the historical customer expression table based on the large language model to obtain the target customer expression; then, the target customer expression is matched with the historical intent table based on the large language model to determine whether there is a historical customer intention that matches the target customer expression; when there is a historical customer intention that matches the target customer expression, the historical customer intention is determined as the target customer intention.

[0129] Optionally, when semantic extraction is performed on the target call text, the historical intent table, and the historical customer expression table based on the large language model, first, the input data is converted into a format that the model can understand, such as a word vector or sentence embedding; then the input data is semantically analyzed using a deep learning algorithm to identify and understand the key information, context, and potential meaning in the text, and finally, based on the understanding of the semantics, the key expression is extracted from the target call text, and the key expression is the target customer expression. Optionally, the target customer expression can be one or more sentences, depending on the complexity of the target call text and the configuration of the large language model.

[0130] It should be noted that since the large language model has the ability to conduct multiple rounds of conversations, in the first round of conversation, the large language model is used to extract the target customer expression from the target call text; in the second round of conversation, the target customer expression is matched with the historical customer intent in the historical intent table, thereby splitting the extraction of the target customer expression and the matching of the target customer intent into two steps. This can reduce the difficulty of the large language model in processing the target call text and ensure the accuracy of the processing results of the large language model.

[0131] Optionally, when matching the target customer expression with the historical intent table based on a large language model to determine whether there is a historical customer intent that matches the target customer expression, first, the target customer expression and the intent statements in the historical intent table are respectively converted into feature vectors; secondly, the similarity between the feature vector of the target customer expression and the feature vector of each intent in the historical intent table is calculated, specifically, it can be achieved by the cosine similarity method; finally, based on the similarity calculation result, it is determined whether it exceeds a preset similarity threshold. If so, it is determined that the historical customer intent matches the target customer expression, wherein the preset similarity threshold is a pre-set value for determining whether two feature vectors are sufficiently similar. The specific setting value of the preset similarity threshold depends on the specific application scenario and requirements, and is not specifically limited in this embodiment.

[0132] Optionally, for the same target customer expression, the generation parameters of the large language model are adjusted to perform multiple repeated reasoning processes in order to understand and parse the target customer expression from multiple angles and levels, so as to capture multiple possible intentions; secondly, during the reasoning process, the historical customer intentions in the historical intention table are sorted in multiple ways, including forward sorting, reverse sorting, and random sorting, which helps to discover the potential correlation and rules of customer intentions under different sortings, so as to more comprehensively understand the real needs of customers; finally, the historical customer intention with the highest degree of consistency in multiple reasoning results or multiple historical customer intentions with higher consistency are counted as the final target customer intention, ensuring that the target customer intentions that match the target customer expression are accurately screened out.

[0133] Optionally, when it is determined that there is a match between the historical customer intention and the target customer statement, the matching relationship between the target customer statement and the historical customer intention is marked. The historical customer intention matches the target customer statement, indicating that the intention expressed by the target customer statement is consistent with the intention in a historical intention table.

[0134] Optionally, the target customer intent is output as a simple text string, a structured JSON object, or other data format suitable for subsequent processing, and stored in a database for subsequent use and analysis.

[0135] Optionally, if a completely matching historical customer intent is not found, further reasoning and analysis can be performed based on the semantic features expressed by the target customer and the relevant information in the historical intent table to try to generate a new intent that is more consistent with the target customer's expression. Alternatively, the initial parameters of the large model or the input order of the target call text can be adjusted to perform matching again based on the large language model.

[0136] The call auxiliary database update method provided in this embodiment inputs the target call text, the historical intent table, and the historical customer expression table corresponding to the historical intent table into the large language model for processing to obtain the target customer expression and the target customer intention corresponding to the target customer expression, including: semantic extraction of the target call text, the historical intent table, and the historical customer expression table based on the large language model to determine the target customer expression; matching the target customer expression with the historical intent table based on the large language model to determine whether there is a historical customer intention that matches the target customer expression; in response to determining that there is a historical customer intention that matches the target customer expression, determining the historical customer intention as the target customer intention. The method of this embodiment can efficiently extract the target customer expression and efficiently match the target customer expression with the historical intention to determine the target customer intention by performing semantic extraction and intention matching on the target call text through the large language model.

[0137] As an optional implementation, based on any of the above embodiments, it also includes: in response to determining that no historical customer intention matches the target customer expression, generating the target customer intention based on the large language model and the target customer expression.

[0138] Specifically, when it is determined that no historical customer intent matches the target customer expression, the large language model will no longer rely on the historical intent table, but will directly generate the target customer intent based on the semantic content of the target customer expression.

[0139] Optionally, when generating the target customer intent from the semantic content of the target customer's expression, firstly, an in-depth semantic analysis is performed on the target customer's expression to understand the key information, context and potential meaning in the target customer's expression, and the core content of the target customer's expression is captured by analyzing the vocabulary, grammatical structure and context; secondly, after completing the semantic analysis, the target customer intent is generated according to the semantic content of the target customer's expression, wherein the large language model may generate the target customer intent that matches the target customer's expression by learning the input historical intent table and historical customer expression table, or by using the large amount of language patterns and knowledge base stored internally. In addition, in the process of generating the target customer intent, the large language model will also perform a certain degree of knowledge reasoning, including using the common sense knowledge, domain knowledge and contextual information in the large language model to infer the target customer's true intent. Through the knowledge reasoning ability, the large language model can more accurately understand the implicit information and potential needs in the target customer's expression, thereby generating a target customer intent that better meets the customer's expectations.

[0140] The call assistance database updating method provided in this embodiment further includes: in response to determining that no historical customer intent matches the target customer expression, generating a target customer intent based on the large language model and the target customer expression. The method of this embodiment can accurately generate the corresponding target customer intent when facing a new or unseen target customer expression, ensuring that the historical customer intent table can be effectively updated.

[0141] As an optional implementation, based on any one of the above embodiments, the target call text includes multiple paragraph contents, and the multiple paragraph contents are obtained by simplifying the historical call text with a large language model; semantic extraction is performed on the target call text, the historical intent table and the historical customer statement table based on the large language model to determine the target customer statement, and it also includes: splicing the multiple paragraph contents with the historical intent table and the historical customer statement table one by one to obtain the text to be input; and inputting the text to be input into the large language model.

[0142] Specifically, in this embodiment, the target call text includes multiple paragraph contents, and the multiple paragraph contents are obtained by simplifying the historical call text with a large language model; before semantic extraction is performed on the target call text, the historical intent table and the historical customer expression table based on the large language model to determine the target customer expression, the multiple paragraph contents are spliced ​​one by one with the historical intent table and the historical customer expression to obtain the text to be input and input into the large language model for processing.

[0143] It should be noted that the purpose of splicing the content of each paragraph of the target call text with the historical intent table and the historical customer expression table one by one is to construct a text to be input that contains complete contextual information so that the large language model can more accurately understand and extract the semantic content. Specifically, when the large language model performs subsequent recognition, it can compare and learn the content of each paragraph with the historical intent table and the historical customer expression table one by one, ensuring the efficiency of the large language model in processing the target call text, and enabling the large language model to deeply and accurately extract the customer expression corresponding to each paragraph, effectively improving the large language model's ability to understand the target call text, ensuring that it can accurately capture the customer's true intentions and extract the target customer's expression.

[0144] Optionally, the splicing method can be a simple text connection or an intelligent splicing based on contextual logic to ensure that the spliced ​​text can fully and accurately reflect the original information.

[0145] The call auxiliary database update method provided in this embodiment, the target call text includes multiple paragraph contents, and the multiple paragraph contents are obtained by simplifying the historical call text with a large language model; semantic extraction is performed on the target call text, the historical intent table, and the historical customer expression table based on the large language model to determine the target customer expression, and it also includes: splicing the multiple paragraph contents with the historical intent table and the historical customer expression table one by one to obtain the text to be input; inputting the text to be input into the large language model. The method of this embodiment, by splicing multiple paragraph contents with the historical intent table and the historical customer expression table, allows the subsequent large language model to compare and learn each paragraph content one by one, thereby accurately extracting the target customer expression corresponding to each paragraph content.

[0146] As an optional implementation, based on any of the above embodiments, the historical customer statement table is updated based on the target customer statement to obtain a target customer statement table, and the historical intention table is updated based on the target customer intention to obtain a target intention table, including: displaying the target customer statement and the target customer intention corresponding to the target customer statement; in response to receiving a user's confirmation operation on the target customer statement through an input device, updating the target customer statement to the historical customer statement table to obtain a target customer statement table; in response to receiving a user's confirmation operation on the target customer intention through an input device, updating the target customer intention to the historical intention table to obtain a target intention table.

[0147] Specifically, when the historical customer expression table is updated based on the target customer expression to obtain the target customer expression table, and the historical intention table is updated based on the target customer intention to obtain the target intention table. First, the target customer expression and the target customer intention corresponding to the target customer expression are displayed so that the user can see the target customer expression and the corresponding target customer intention; secondly, when the user confirms the target customer expression through the input device, the target customer expression is updated to the historical customer expression table; when the user confirms the target customer intention through the input device, the target customer intention is updated to the historical intention table, thereby obtaining the target customer expression table and the target intention table.

[0148] It should be noted that by displaying the target customer statements and target customer intentions, users can view the target customer statements and target customer intentions. As a result, users can judge the target customer statements and target customer intentions based on their own professional knowledge and experience. Only when users confirm that they are new and valuable customer statements and customer intentions will they perform confirmation operations, thereby avoiding adding duplicate, invalid or erroneous information to the call assistance database, and ensuring that the call assistance database is updated accurately and effectively.

[0149] Optionally, the target customer expression obtained by processing the large language model and its corresponding target customer intention are displayed to the user through a display device. Optionally, the target customer expression and the target customer intention can be displayed on the same user interface and presented in a list, table or other form. Optionally, the display device can be a computer monitor, a mobile phone screen, a tablet computer, etc.

[0150] Optionally, the input device may be a mouse, a keyboard or a touch screen, etc. The user's confirmation operation on the displayed target customer statement may be a click, a selection or an input command, etc.

[0151] Optionally, updating the target customer expression to the historical customer expression table and updating the target customer intention to the historical intention table may include: adding a new entry, modifying an existing entry, or deleting an entry.

[0152] Optionally, the content in the target call text corresponding to the target customer's statement is displayed, so that when the user checks the target customer's statement and the target customer's intention, they can confirm whether the target customer's statement matches the content in the target call text through the content of the displayed target call text, thereby ensuring the accuracy of the target customer's statement.

[0153] The call assistance database update method provided in this embodiment updates the historical customer expression table based on the target customer expression to obtain the target customer expression table, and updates the historical intention table based on the target customer intention to obtain the target intention table, including: displaying the target customer expression and the target customer intention corresponding to the target customer expression; in response to receiving the user's confirmation operation on the target customer expression through the input device, updating the target customer expression to the historical customer expression table to obtain the target customer expression table; in response to receiving the user's confirmation operation on the target customer intention through the input device, updating the target customer intention to the historical intention table to obtain the target intention table. The method of this embodiment displays the target customer expression and the target customer intention for the user to review and confirm. Only after receiving the user's confirmation operation, the corresponding target customer expression and target customer intention are updated, ensuring that the call assistance database can be updated accurately and effectively.

[0154] As an optional implementation manner, based on any of the above embodiments, it also includes: sorting the target customer expressions according to a preset sorting rule; and displaying the sorted target customer expressions and the target customer intentions corresponding to the target customer expressions.

[0155] Specifically, before displaying the target customer statements and target customer intentions, first, the target customer statements are sorted according to a preset sorting rule, and then the sorted target customer statements and the target customer intentions corresponding to the target customer statements are displayed, so as to facilitate the user to view and confirm.

[0156] The preset sorting rule is a pre-set rule for sorting the target customer statements. The preset sorting rule can be formulated according to actual needs, for example, sorting according to the frequency of occurrence, importance, time sequence or semantic relevance of the target customer statements. In this embodiment, the preset sorting rule is not specifically limited.

[0157] It should be noted that the extracted target customer statements are sorted according to the preset sorting rules. The sorted target customer statements will be arranged in a specific order so that users can review and confirm the important contents first.

[0158] The call assistance database update method provided in this embodiment also includes: sorting the target customer expressions according to a preset sorting rule; and displaying the sorted target customer expressions and the target customer intentions corresponding to the target customer expressions. The method of this embodiment can optimize the order in which users view the target customer expressions by sorting the target customer expressions according to the preset sorting rule before displaying them, so that users can confirm the target customer expressions and target customer intentions with higher relevance or importance, thereby improving the update efficiency of the call assistance database.

[0159] As an optional implementation, based on any of the above embodiments, the target customer expressions are sorted according to a preset sorting rule, including: sorting the target customer expressions according to semantic similarity, repetition frequency and meaning. Figure 1 Any one or more combinations of the degree of consistency are used to sort.

[0160] Specifically, when the target customer expressions are sorted according to the preset sorting rules, the target customer expressions are sorted according to the semantic similarity, repetition frequency and meaning. Figure 1 Thus, the target customer statements can be sorted from multiple angles to obtain more accurate and meaningful sorting results, thereby ensuring that users give priority to reviewing and confirming target customer statements with high maintenance value.

[0161] Semantic similarity refers to the semantic similarity between the target customer's expression and the target customer's intention, and is an important indicator for measuring the degree of match between the target customer's expression and the target customer's intention. Optionally, the word vector distance between the target customer's expression and the target customer's intention can be calculated by a word vector model (such as word2vec); or, the semantic similarity can be calculated using cosine similarity or Jaccard similarity.

[0162] The repetition frequency refers to the number of times any target customer statement appears in all target customer statements. A target customer statement with a higher frequency of appearance may be more representative or important, and thus may be displayed first.

[0163] Optionally, before counting the repetition frequency of target customer expressions, all target customer expressions are first standardized, that is, syntactic component analysis is performed to identify the subject, predicate and object in the sentence. On the basis of syntactic analysis, only the core components such as the subject, predicate and object in the sentence are retained to eliminate redundant information in the expression, thereby ensuring that even if the customer uses different words or sentence patterns to express the same meaning, they can be correctly identified as the same expression after standardization.

[0164] Among them, Figure 1 The degree of consistency refers to the degree of match between the target customer's expression and the corresponding target customer's intention when the large language model is used to match the target customer's expression with the historical intention. It should be noted that due to the diversity and complexity of customer expressions, a single match may not be able to fully capture the customer's true intention. Therefore, multiple repeated matches are performed through the large language model to analyze the customer's expression from multiple angles and levels in order to obtain more comprehensive and accurate matching results. In multiple matching results, the large language model will sort the results according to the degree of consistency between the target customer's expression and the matched target customer's intention. From the sorted matching results, the customer intention with the highest degree of consistency or the top three customer intentions with a high degree of consistency are selected for output. Optionally, by adjusting the large language model inference parameters (such as generation temperature); or, the historical customer intentions in the historical customer intention table are sorted forward, reversely, or randomly, so that the large language model can be repeated for inference multiple times.

[0165] Optionally, after the target customer expressions are sorted by semantic similarity, repetition frequency and meaning, Figure 1 When combining and sorting according to the degree of consistency, the weight of each sorting rule can be set according to actual needs, and the final sorting result can be obtained by weighted summation or merging after sorting. Specifically, in this embodiment, the specific weight value of each sorting rule is not limited, and can be flexibly set according to actual needs.

[0166] The call assistance database updating method provided in this embodiment sorts the target customer expressions according to a preset sorting rule, including: sorting the target customer expressions according to semantic similarity, repetition frequency and meaning. Figure 1 The method of this embodiment can sort the target customer statements from multiple angles to obtain more accurate and meaningful sorting results, thereby ensuring that users give priority to reviewing and confirming target customer statements with high maintenance value.

[0167] As an optional implementation, based on any one of the above embodiments, it also includes: in response to determining that the text length of the historical call text is less than or equal to a preset text length, the historical call text, the historical intent table, and the historical customer statement table corresponding to the historical intent table are input into the large language model for semantic extraction and intent matching to obtain the target customer statement and the target customer intent corresponding to the target customer statement.

[0168] Specifically, when it is determined that the text length of the historical call text is less than or equal to the preset text length, the historical call text, the historical intent table, and the historical customer statement table corresponding to the historical intent table are input into the large language model for semantic extraction and intent matching to obtain the target customer statement and the target customer intent corresponding to the target customer statement.

[0169] It should be noted that before processing the historical call text, it is first determined whether the length of the historical text is less than or equal to a preset text length threshold. The preset text length can be set according to actual needs, for example, it can be several thousand words, tens of thousands of words or longer.

[0170] It should be noted that if the text length of the historical call text is less than or equal to the preset text length, then there is no need to simplify it. It can be directly concatenated with the historical intent table and the corresponding historical customer expression table and input into the large language model together.

[0171] Optionally, after receiving the input, the large language model will perform semantic extraction, that is, analyze and understand the semantic information in the text and extract the target customer expression. On this basis, the large language model will also perform intent matching, that is, match the extracted target customer expression with the historical intent in the historical intent table to find the target customer intent that matches the target customer expression.

[0172] The call assistance database update method provided in this embodiment also includes: in response to determining that the text length of the historical call text is less than or equal to the preset text length, inputting the historical call text, the historical intent table, and the historical customer expression table corresponding to the historical intent table into the large language model for semantic extraction and intent matching to obtain the target customer expression and the target customer intention corresponding to the target customer expression. The method of this embodiment, for historical call texts whose text length is less than or equal to the preset text length, does not need to be simplified, and directly inputs the historical call text, the historical intent table, and the historical customer expression table into the large language model for processing to obtain the target customer expression and the target customer intention, thereby ensuring efficient updating of the call assistance database.

[0173] Figure 3 A flowchart of a method for updating a call assistance database provided by another embodiment of the present application is shown below. Figure 3 As shown in , this embodiment provides a complete process for updating a call assistance database. Specifically, the call assistance database updating method of this embodiment includes:

[0174] S301. Obtain historical call texts, historical intent tables, and historical customer description tables.

[0175] Specifically, the historical call text is converted from the historical call record voice through a voice-to-text tool. Among them, the historical intention table and the historical customer expression table are both obtained from the call auxiliary database.

[0176] S302: Select a resource configuration mode for a large language model.

[0177] Specifically, the task of obtaining target customer expressions and corresponding target customer intentions is divided into five different reasoning modules according to the reasoning difficulty, including:

[0178] 1) Paragraph splitting (simple reasoning, model parameters need to be at the level of tens of billions or seventy billion)

[0179] 2) Paragraph summary (simple reasoning, model parameters need to be at the level of 10 billion or 70 billion)

[0180] 3) Target customer description extraction (simple reasoning, model parameters require tens of billions or seventy billion levels)

[0181] 4) Target customer intent matching (complex reasoning, model parameters require 70 billion levels)

[0182] 5) Generate new customer intent (complex reasoning, model parameters require 70 billion level)

[0183] It should be noted that when the parameters of a large language model need to be at the level of tens of billions, it means that the large language model has about 10^10 (i.e., 10 billion) trainable parameters. These parameters include weights and biases in the neural network, which are adjusted through the back-propagation algorithm during the model training process so that the model can learn the rules in the input data and make accurate predictions or generate reasonable outputs. The number of parameters at the level of tens of billions makes the large language model have higher complexity and expressiveness, and can handle more complex and diverse natural language tasks. When the number of parameters reaches the level of 70 billion, that is, about 7×10^10 (i.e., 70 billion) trainable parameters, this indicates that the model is larger in scale and more complex. Such a model theoretically has stronger expressiveness and the ability to learn complex patterns. The number of parameters at the level of 70 billion enables the model to have a deeper understanding of the structure, grammar, word meaning, and contextual relationships of natural language, thereby showing higher accuracy in natural language processing tasks.

[0184] Therefore, this embodiment has two optional resource configuration modes for large language models. One is to configure model parameters of 70 billion levels for all five reasoning modules under the condition that data processing resources allow; the other is to configure model parameters of 10 billion levels for modules with low reasoning difficulty and low requirements for model parameters, such as paragraph splitting, paragraph summary and target customer expression extraction, under the condition that data processing resources are limited. For modules with high reasoning difficulty and higher requirements for model parameters and computing power, such as target customer intent matching and new customer intent generation, model parameters of 70 billion levels are configured to ensure the accuracy of data processing. Specifically, which resource configuration mode to choose can be selected according to actual conditions.

[0185] S303: Determine whether the historical call text is an extra-long text.

[0186] Specifically, the historical call text is compared with the preset text length to confirm whether it is an overlong text.

[0187] Specifically, when it is determined to be a short text, step S304 is executed; when it is determined to be a very long text, step S305 is executed.

[0188] S304: When it is determined that the historical call text is a short text, the historical call text, the historical intent table and the historical customer expression table are input into the large language model for semantic extraction and intent matching to obtain the target customer expression and the target customer intent.

[0189] Specifically, when the text length of the historical call text is less than or equal to the preset text length, it is determined to be a short text.

[0190] Specifically, after obtaining the target customer's expression and the target customer's intention, step S314 is executed.

[0191] S305: When it is determined that the historical call text is an overlong text, the historical call text is split based on the large language model and preset prompt words to obtain multiple initial paragraphs.

[0192] S306: Based on the large language model, the contents of the multiple initial paragraphs are refined and summarized to generate multiple summaries.

[0193] S307: Merge the content of the current initial paragraph with the summaries corresponding to other initial paragraphs to obtain a target paragraph.

[0194] S308: Merge multiple target paragraphs to obtain a target call text.

[0195] S309: Connect the contents of multiple paragraphs with the historical intent table and the historical customer expression table one by one to obtain the text to be input, and input the text to be input into the large language model.

[0196] S310: Perform semantic extraction on the input text based on the large language model to determine the target customer expression.

[0197] S311. Match the target customer's expression with the historical intent table based on the large language model to determine whether there is a historical customer intent that matches the target customer's expression.

[0198] Specifically, when it is determined that the historical customer intention matches the target customer statement, step S312 is executed; when it is determined that the historical customer intention does not match the target customer statement, step S313 is executed.

[0199] S312. When there is a match between the historical customer intention and the target customer expression, the historical customer intention is determined as the target customer intention.

[0200] Specifically, when the historical customer intention is determined to be the target customer intention, step S314 is executed.

[0201] S313: When no historical customer intent matches the target customer expression, the target customer intent is generated based on the large language model and the target customer expression.

[0202] S314, sort the target customer statements according to semantic similarity, repetition frequency and meaning Figure 1 Any one or more combinations of the degree of consistency are used to sort.

[0203] S315: Display the sorted target customer statements and the target customer intentions corresponding to the target customer statements.

[0204] Specifically, after the target customer statement and target customer intention are displayed, the auditor can review the target customer statement and target customer intention and conduct an audit and confirmation based on his or her own professional knowledge.

[0205] Optionally, in addition to displaying the target customer statement and the target customer intention, the target paragraph corresponding to the target customer statement will also be displayed to assist the auditor in conducting the audit confirmation.

[0206] S316. When a confirmation operation of the target customer expression by the user is received through the input device, the target customer expression is updated to the historical customer expression table; when a confirmation operation of the target customer intention by the user is received through the input device, the target customer intention is updated to the historical intention table.

[0207] Specifically, the auditor determines the correct and effective target customer representation and target customer intention through the input device.

[0208] Specifically, the updated historical intention table and historical customer representation table are sent to the call assistance database to replace the original historical intention table and historical customer representation table.

[0209] Figure 4 A schematic diagram of the structure of a call assistance database updating device provided in an embodiment of the present application is shown in FIG. Figure 4 As shown, the communication assistance database updating device provided in this embodiment is located in an electronic device. The communication assistance database updating device 40 provided in this embodiment includes: an acquisition module 41, a simplification module 42, a processing module 43 and an updating module 44.

[0210] Specifically, an acquisition module 41 is used to acquire historical call texts; a simplification module 42 is used to, in response to determining that the text length of the historical call text is greater than a preset text length, input the historical call text into a large language model for simplification processing to obtain a target call text; a processing module 43 is used to input the target call text, a historical intent table, and a historical customer expression table corresponding to the historical intent table into a large language model for processing to obtain a target customer expression and a target customer intent corresponding to the target customer expression; an update module 44 is used to update the historical customer expression table based on the target customer expression to obtain a target customer expression table, and to update the historical intent table based on the target customer intent to obtain a target intent table.

[0211] Optionally, when the historical call text is input into the large language model for simplification processing to obtain the target call text, the simplification module 42 is specifically used to: split the historical call text based on the large language model and preset prompt words to obtain multiple initial paragraphs; refine and summarize the contents of the multiple initial paragraphs based on the large language model to generate multiple summaries; merge the contents of the multiple initial paragraphs with the multiple summaries based on preset merging rules to obtain the target call text.

[0212] Optionally, the processing module 43, when inputting the target call text, the historical intent table and the historical customer expression table corresponding to the historical intent table into the large language model for processing to obtain the target customer expression and the target customer intention corresponding to the target customer expression, is specifically used to: perform semantic extraction on the target call text, the historical intent table and the historical customer expression table based on the large language model to determine the target customer expression; match the target customer expression with the historical intent table based on the large language model to determine whether there is a historical customer intention that matches the target customer expression; in response to determining that there is a historical customer intention that matches the target customer expression, determine the historical customer intention as the target customer intention.

[0213] Optionally, the processing module 43 is further configured to: in response to determining that no historical customer intent matches the target customer expression, generate a target customer intent based on the large language model and the target customer expression.

[0214] Optionally, the target call text includes multiple paragraph contents, and the multiple paragraph contents are obtained by simplifying the historical call text with the large language model; the processing module 43, before performing semantic extraction on the target call text, the historical intent table and the historical customer expression table based on the large language model to determine the target customer expression, is also used to: splice the multiple paragraph contents with the historical intent table and the historical customer expression table one by one to obtain the text to be input; and input the text to be input into the large language model.

[0215] Optionally, the update module 44, when updating the historical customer statement table based on the target customer statement to obtain the target customer statement table, and updating the historical intention table based on the target customer intention to obtain the target intention table, is specifically used to: display the target customer statement and the target customer intention corresponding to the target customer statement; in response to receiving a confirmation operation of the target customer statement by the user through an input device, update the target customer statement to the historical customer statement table to obtain the target customer statement table; in response to receiving a confirmation operation of the target customer intention by the user through an input device, update the target customer intention to the historical intention table to obtain the target intention table.

[0216] Optionally, the updating module 44 is further configured to: sort the target customer expressions according to a preset sorting rule; and display the sorted target customer expressions and the target customer intentions corresponding to the target customer expressions.

[0217] Optionally, the updating module 44, when sorting the target customer expressions according to the preset sorting rules, is specifically used to: sort the target customer expressions according to the semantic similarity, repetition frequency and meaning. Figure 1 Any one or more combinations of the degree of consistency are used to sort.

[0218] Optionally, the processing module 43 is also used to: in response to determining that the text length of the historical call text is less than or equal to a preset text length, input the historical call text, the historical intent table, and the historical customer statement table corresponding to the historical intent table into the large language model for semantic extraction and intent matching to obtain the target customer statement and the target customer intent corresponding to the target customer statement.

[0219] Figure 5 A schematic diagram of the structure of an electronic device provided in an embodiment of the present application is shown in FIG. Figure 5 As shown, the electronic device 50 provided in this embodiment includes: a processor 51 and a memory 52 communicatively connected to the processor.

[0220] The memory 52 stores computer-executable instructions; the processor 51 executes the computer-executable instructions stored in the memory 52 to implement the call assistance database updating method provided in any one of the above embodiments.

[0221] The program may include program codes, and the program codes include computer-executable instructions. The memory 52 may include a high-speed RAM memory, and may also include a non-volatile memory (non-volatile memory), such as at least one disk memory.

[0222] In this embodiment, the memory 52 is connected to the processor 51 via a bus. The bus may be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus. The bus may be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 5 The fact that only one straight line is used in the diagram does not mean that there is only one bus or only one type of bus.

[0223] An embodiment of the present application also provides a computer-readable storage medium, including computer-executable instructions stored in the computer-readable storage medium, and when the computer-executable instructions are executed by a processor, they are used to implement the call assistance database update method provided in any of the above embodiments.

[0224] An embodiment of the present application also provides a computer program product, including a computer program, which, when executed by a processor, implements the call assistance database update method provided in any one of the above embodiments.

[0225] The modules described as separate components may or may not be physically separated, and the components shown as modules may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the modules may be selected according to actual needs to implement the solution of this embodiment.

[0226] In addition, each functional module in each embodiment of the present application can be integrated into one processing unit, or each module can exist physically separately, or two or more modules can be integrated into one unit. The above-mentioned module-composed unit can be implemented in the form of hardware or in the form of hardware plus software functional units.

[0227] The above-mentioned integrated module implemented in the form of a software function module can be stored in a computer-readable storage medium. The above-mentioned software function module is stored in a storage medium, including a number of instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) or a processor to perform some steps of the methods of various embodiments of the present application.

[0228] It should be understood that the above processor can be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), etc. A general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in the invention can be directly embodied as being executed by a hardware processor, or can be executed by a combination of hardware and software modules in the processor.

[0229] The memory may include a high-speed RAM memory, and may also include a non-volatile storage NVM, such as at least one disk memory, and may also be a USB flash drive, a mobile hard disk, a read-only memory, a magnetic disk or an optical disk, etc.

[0230] The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. The bus can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, the bus in the drawings of this application is not limited to only one bus or one type of bus.

[0231] The above storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk or optical disk. The storage medium can be any available medium that can be accessed by a general or special purpose computer.

[0232] An exemplary storage medium is coupled to a processor so that the processor can read information from the storage medium and write information to the storage medium. Of course, the storage medium can also be a component of the processor. The processor and the storage medium can be located in an application specific integrated circuit (ASIC). Of course, the processor and the storage medium can also exist as discrete components in an electronic control unit or a main control device.

[0233] Those skilled in the art can understand that all or part of the steps of implementing the above-mentioned method embodiments can be completed by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, the steps of the above-mentioned method embodiments are executed; and the aforementioned storage medium includes: ROM, RAM, disk or optical disk and other media that can store program codes.

[0234] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit it. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for updating a call assistance database, characterized in that: include: Get historical call texts; In response to determining that the text length of the historical call text is greater than a preset text length, inputting the historical call text into a large language model for simplification processing to obtain a target call text; Inputting the target call text, the historical intent table, and the historical customer expression table corresponding to the historical intent table into a large language model for processing to obtain a target customer expression and a target customer intention corresponding to the target customer expression; The historical customer expression table is updated based on the target customer expression to obtain a target customer expression table, and the historical intent table is updated based on the target customer intent to obtain a target intent table.

2. The method according to claim 1, characterized in that The step of inputting the historical call text into the large language model for simplification processing to obtain the target call text includes: Splitting the historical call text based on the large language model and preset prompt words to obtain multiple initial paragraphs; Based on the large language model, the contents of multiple initial paragraphs are refined and summarized to generate multiple summaries; The contents of the multiple initial paragraphs are merged with the multiple summaries based on a preset merging rule to obtain a target call text.

3. The method according to claim 2, characterized in that The step of merging the contents of the plurality of initial paragraphs with the plurality of summaries based on a preset merging rule to obtain a target call text includes: Merge the content of the current initial paragraph with the summaries corresponding to other initial paragraphs to obtain a target paragraph; Merge multiple target paragraphs to obtain the target call text.

4. The method according to claim 1, characterized in that: The step of inputting the target call text, the historical intent table, and the historical customer expression table corresponding to the historical intent table into the large language model for processing to obtain the target customer expression and the target customer intention corresponding to the target customer expression includes: Performing semantic extraction on the target call text, the historical intent table, and the historical customer expression table based on a large language model to determine the target customer expression; Matching the target customer expression with the historical intent table based on a large language model to determine whether there is a historical customer intent that matches the target customer expression; In response to determining that there is a historical customer intent that matches the target customer expression, the historical customer intent is determined as the target customer intent.

5. The method according to claim 4, characterized in that Also includes: In response to determining that no historical customer intent matches the target customer expression, a target customer intent is generated based on a large language model and the target customer expression.

6. The method according to claim 4, characterized in that The target call text includes a plurality of paragraphs, and the plurality of paragraphs are obtained by simplifying the historical call text with a large language model; Before the semantic extraction of the target call text, the historical intent table and the historical customer expression table based on the large language model to determine the target customer expression, the method further includes: splicing the contents of multiple paragraphs with the historical intent table and the historical customer expression table one by one to obtain the text to be input; The text to be input is input into the large language model.

7. The method according to claim 1, characterized in that The updating of the historical customer expression table based on the target customer expression to obtain a target customer expression table, and the updating of the historical intention table based on the target customer intention to obtain a target intention table, include: Displaying the target customer statement and the target customer intention corresponding to the target customer statement; In response to receiving a confirmation operation of the user on the target customer statement through an input device, updating the target customer statement to the historical customer statement table to obtain a target customer statement table; In response to receiving a determination operation of the target customer intention by the user through an input device, the target customer intention is updated to the historical intention table to obtain a target intention table.

8. The method according to claim 7, characterized in that Also includes: Sorting the target customer statements according to a preset sorting rule; The sorted target customer expressions and the target customer intentions corresponding to the target customer expressions are displayed.

9. The method according to claim 8, characterized in that Sorting the target customer statements according to preset sorting rules, including: The target customer expressions are sorted according to any one or a combination of semantic similarity, repetition frequency, and consistency of intention.

10. The method according to any one of claims 1 to 9, characterized in that Also includes: In response to determining that the text length of the historical call text is less than or equal to a preset text length, the historical call text, the historical intent table, and the historical customer statement table corresponding to the historical intent table are input into a large language model for semantic extraction and intent matching to obtain a target customer statement and a target customer intent corresponding to the target customer statement.

11. A call assistance database updating device, characterized in that: include: The acquisition module is used to obtain historical call texts; a simplification module, configured to, in response to determining that the text length of the historical call text is greater than a preset text length, input the historical call text into a large language model for simplification processing to obtain a target call text; A processing module, configured to input the target call text, the historical intent table, and the historical customer expression table corresponding to the historical intent table into a large language model for processing, so as to obtain a target customer expression and a target customer intent corresponding to the target customer expression; An updating module is used to update the historical customer expression table based on the target customer expression to obtain a target customer expression table, and to update the historical intent table based on the target customer intent to obtain a target intent table.

12. An electronic device, characterized in that: include: A processor, and a memory communicatively connected to the processor; The memory stores computer-executable instructions; The processor executes the computer-executable instructions stored in the memory to implement the method according to any one of claims 1 to 10.

13. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer-executable instructions, which are used to implement the method according to any one of claims 1 to 10 when executed by a processor.

14. A computer program product, characterized in that The invention comprises a computer program, which implements the method according to any one of claims 1 to 10 when being executed by a processor.

Images