Abstract generation method, device and platform and storage medium

Abstract

The invention provides an abstract generation method and device, a platform and a storage medium. The method comprises the following steps: on the basis of a preset mapping relationship between message formats and processing logics, respectively searching for the processing logics required by a plurality of chat messages; according to the preset content dimension in the processing logic searched for the chat message, the message content corresponding to the required content dimension is extracted from the chat message, and the length of the message content is smaller than that of the chat message; and constructing abstract data corresponding to the chat record by using the processing data output by the target language model. Therefore, redundancy elimination processing can be performed on the chat message based on the mapping relation provided by the embodiment of the invention to compress the length of the chat message, so that the reading time of the whole chat message is saved, the automation of abstract generation can be realized by means of the target language model, the time of repeatedly reading the chat record is further saved, and the user experience is improved. And the generation efficiency of the abstract is effectively improved.

Description

Technical Field

[0001] The present disclosure relates to the field of Internet technology, and in particular to a summary generation method, device, platform and storage medium. Background Art

[0002] Currently, in the scenario of home services, when a customer has a demand for home services, the customer can publish the demand information through the housekeeping platform, and the housekeeping platform will send the demand information to the broker. After the broker finds a housekeeper who matches the demand information, the broker will feedback the housekeeper's identity information to the customer through the housekeeping platform. The customer can continuously negotiate with the broker about the required housekeeper. In this way, there will be a large number of chat records between the broker and the customer. Since the chat content recorded in the chat records is complicated, it is necessary to summarize the chat records so as to use the summary to quickly understand the chat content.

[0003] Existing summary generation solutions usually adopt the method of manually generating summaries of chat records. Since chat records contain a large number of chat messages, the chat content to be expressed in each chat message may be different. Therefore, this method requires manual reading of each chat message in the chat record multiple times, and after understanding the chat content corresponding to each chat message, a summary of the chat record is generated. This results in a lot of time consumption and affects the efficiency of summary generation. Summary of the invention

[0004] Multiple aspects of the present application provide a summary generation method, device, platform and storage medium to improve the efficiency of summary generation.

[0005] The present application provides a method for generating a summary, including:

[0006] Obtaining a chat record to be processed, wherein the chat record includes multiple chat messages;

[0007] Determining message formats corresponding to each of the plurality of chat messages;

[0008] Based on the mapping relationship between the preset message format and the processing logic, searching for the processing logic required for the multiple chat messages respectively;

[0009] For any chat message, according to the content dimension preset in the processing logic found for the chat message, extract the message content corresponding to the required content dimension from the chat message, wherein the length of the message content is less than the length of the chat message;

[0010] Inputting the message contents generated for the plurality of chat messages respectively as input data into the target language model;

[0011] Inputting preset prompt information into the target language model, wherein the prompt information includes task description information for instructing to generate a summary based on the input data;

[0012] Using the target language model to perform natural language understanding on the input data according to the task description information to obtain processed data;

[0013] Based on the processed data, summary data corresponding to the chat record is constructed.

[0014] The present application also provides a summary generation device, including:

[0015] A data acquisition unit, used to acquire a chat record to be processed, wherein the chat record includes a plurality of chat messages;

[0016] A data determination unit, used to determine the message formats corresponding to the plurality of chat messages;

[0017] A logic search unit, used to search for the processing logic required for the plurality of chat messages respectively based on a mapping relationship between a preset message format and a processing logic;

[0018] a logic processing unit, configured to extract, for any chat message, message content corresponding to a required content dimension from the chat message according to a content dimension preset in a processing logic found for the chat message, wherein the length of the message content is less than the length of the chat message;

[0019] A first data input unit, configured to input the message contents generated for the plurality of chat messages respectively as input data into a target language model;

[0020] A second data input unit, configured to input preset prompt information into the target language model, wherein the prompt information includes task description information for instructing to generate a summary based on the input data;

[0021] A data processing unit, configured to perform natural language understanding on the input data according to the task description information using the target language model to obtain processed data;

[0022] A summary construction unit is used to construct summary data corresponding to the chat record based on the processed data.

[0023] The embodiment of the present application also provides a computing platform, including a memory and a processor;

[0024] The memory is used to store one or more computer instructions;

[0025] The processor is coupled to the memory and configured to execute the one or more computer instructions to perform the aforementioned summary generation method.

[0026] The embodiment of the present application also provides a computer-readable storage medium, which, when the computer instructions are executed by one or more processors, causes the one or more processors to execute the aforementioned summary generation method.

[0027] In an embodiment of the present application, a chat record to be processed can be obtained, wherein the chat record contains multiple chat messages; the message format corresponding to each of the multiple chat messages is determined; based on the mapping relationship between the preset message format and the processing logic, the processing logic required for the multiple chat messages is searched respectively; for any chat message, according to the content dimension preset in the processing logic found for the chat message, the message content corresponding to the required content dimension is extracted from the chat message, wherein the length of the message content is less than the length of the chat message; the message content generated for the multiple chat messages is used as input data and input into the target language model; the preset prompt information is input into the target language model, wherein the prompt information contains task description information for indicating the generation of a summary based on the input data; the target language model is used to perform natural language understanding on the input data according to the task description information to obtain processing data; based on the processing data, the summary data corresponding to the chat record is constructed. Accordingly, based on the mapping relationship proposed in the embodiment of the present application, the chat message can be de-redundantly processed to compress the length of the chat message. In this way, when it is necessary to generate a summary for the chat record, a summary can be generated based on the compressed chat content, thereby saving the time of reading the entire chat message. Moreover, the summary of the chat record can be generated with the help of the target language model to realize the automation of summary generation, further saving the time of repeatedly reading the chat record, which can effectively improve the efficiency of summary generation. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0029] Figure 1 A schematic diagram of the structure of a summary generation system provided by an exemplary embodiment of the present application;

[0030] Figure 2 A schematic diagram of an optional implementation logic of a summary generation solution provided by an exemplary embodiment of the present application;

[0031] Figure 3 A flowchart of a method for generating a summary provided by another exemplary embodiment of the present application;

[0032] Figure 4 A schematic diagram of the structure of a summary generation device provided by another exemplary embodiment of the present application;

[0033] Figure 5 A structural diagram of a computing platform provided for yet another exemplary embodiment of the present application. DETAILED DESCRIPTION

[0034] In order to make the purpose, technical solution and advantages of the present application clearer, the technical solution of the present application will be clearly and completely described below in combination with the specific embodiments of the present application and the corresponding drawings. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present application.

[0035] Before starting to explain in detail the technical solutions provided by the embodiments of the present application, several technical concepts involved in the present application are briefly explained as follows.

[0036] The target language model can be understood as a large language model (LLM) with semantic recognition function. In actual application, the large language model can be pre-trained through training samples to learn general language representation and knowledge, and then the large language model can be fine-tuned so that the large language model can be applied to different processing tasks. For example, the large language model can be widely used in the field of natural language processing (NLP), and can be specifically applied to processing tasks in the field of natural language processing such as text-based sentiment classification, text summary generation, and machine translation. The large language model can perform various processing tasks based on the prompt information, and with the help of the Transformer architecture and attention mechanism, it can understand the dependencies between texts and focus on the part of the input data related to the processing task, so as to better understand and output the required content.

[0037] Prompt information is usually used to inform the model of the tasks to be performed and guide the model to output the corresponding content according to the tasks. When the model is required to process text, prompt information can provide the usage scenario of the text and the adjacent text of the text to help the model output the content accurately. Prompt information can also be used to control the length of the model output, thereby improving the refinement of the output content.

[0038] Natural Language Understanding (NLU) can be understood as a general term for methods, models or tasks in the field of artificial intelligence and natural language processing that support computers to understand natural language content, such as the meaning, intent and structure of natural language, so that computers can understand the meaning presented by natural language text and realize natural language communication between humans and computers.

[0039] As introduced in the background technology, in traditional summary generation schemes, a method of manually generating summaries of chat records is usually adopted. Since chat records contain a large number of chat messages, the chat content to be expressed in each chat message may be different. Therefore, this method requires manual reading of each chat message in the chat record for multiple times, and after understanding the chat content corresponding to each chat message, a summary of the chat record is generated. This results in a lot of time consumption and affects the efficiency of summary generation.

[0040] To this end, this embodiment proposes a summary generation solution in the hope of improving the efficiency of summary generation.

[0041] The following describes in detail the various embodiments of the present application in conjunction with the accompanying drawings.

[0042] Figure 1 A schematic diagram of a summary generation system provided by an exemplary embodiment of the present application is shown in FIG. Figure 1 The system may include a summary generating device and a target language model.

[0043] Optionally, the summary generation device in this embodiment can be implemented as a functional plug-in and pluggably connected to the target language model to form a summary generation system. Of course, the summary generation device in this embodiment can also be implemented as other product forms, and the product form of the summary generation device is not limited here.

[0044] In this embodiment, multiple chat samples may be prepared in advance, each of which contains at least two chat roles. For example, in the scenario of home service, the chat roles may be a broker, a customer, and a housekeeper. In this embodiment, the chat samples that meet the summary generation conditions may be described as chat records to be processed.

[0045] In this embodiment, the summary generation condition can be set as that the chat record contains at least N chat messages, and no chat role has sent a chat message within M consecutive time. The summary generation condition can be adaptively modified and adjusted according to the actual application scenario, and the condition parameters (M and N) in the summary generation condition can also be adaptively configured according to the actual application scenario. This embodiment does not specifically limit the content of the summary generation condition and the condition parameters included.

[0046] Figure 2This is a schematic diagram of an optional implementation logic of a summary generation solution provided by an exemplary embodiment of the present application. Figure 2 After obtaining the chat records to be processed (i.e., meeting the summary generation conditions), the chat records can be cleaned.

[0047] Data cleaning can be understood as filtering useless information, duplicate information, and incomplete information in the chat records when preprocessing the chat records. Optionally, in this embodiment, data cleaning may include filtering invalid messages and filtering withdrawn messages, and data cleaning steps may be added according to actual needs, for example, steps of filtering single chat records and filtering target chat records corresponding to chat roles may be added.

[0048] Among them, invalid messages include test messages and dirty messages. Test messages refer to chat messages generated when logging in with a predetermined test account; dirty messages refer to chat messages containing garbled characters, meaningless character sequences, or malicious behaviors. For example, when a chat begins, a chat identifier bound to the chat is generated, and the chat duration of the chat is recorded. Usually, there is a chat identifier (i.e., the bound chat identifier) ​​within the chat duration. If there are multiple chat identifiers within the same chat duration, each chat identifier can be matched with the bound chat identifier, thereby filtering out chat information corresponding to the unmatched chat identifier as a dirty message. Dirty messages can also be filtered out in other ways, for example, detecting the source of each chat identifier and the integrity of the chat message corresponding to each chat identifier, and this embodiment does not provide more examples for this.

[0049] Screening single chat records can be understood as screening out chat records that contain two chat roles. For example, in the scenario of home services, single chat records can be chat records between brokers and customers, chat records between brokers and housekeepers, and chat records between brokers and brokers. Single chat records can also be further divided according to actual needs to obtain external single chat records and internal single chat records. External single chat records are the chat records between brokers and customers and chat records between brokers and housekeepers, and internal single chat records are the chat records between brokers and brokers, so that summary data of the single chat records can be generated later.

[0050] Filtering the target chat records corresponding to the chat roles can be understood as filtering out the chat records containing the specific chat roles. For example, the broker can be determined as the specific chat role, and the chat records corresponding to the broker can be filtered out from the chat records as the target chat records, so as to generate summary data of the target chat records.

[0051] In this embodiment, the chat record contains multiple chat messages, each chat message corresponds to a message format, and the summary generation device can determine the message format corresponding to each of the multiple messages. Here, the message format of the chat message is not limited, and will be exemplified in the following text.

[0052] refer to Figure 1 In this embodiment, the summary generation device can search for the processing logic required for multiple chat messages based on the mapping relationship between the preset message format and the processing logic.

[0053] According to the above description, the chat log contains multiple message formats. Since different message formats usually have different data structures, the data contents contained in different data structures are also different. Therefore, chat messages in different message formats may contain different data contents. In this way, if it is necessary to centrally process multiple message formats with unified processing logic, there will be a problem that a chat message lacks certain data content, resulting in the inability to process the chat message. Therefore, it is necessary to set corresponding processing logic for each message format.

[0054] In this regard, this embodiment proposes a mapping table that represents the mapping relationship between message formats and processing logic, so that the processing logic corresponding to each message format can be found through the mapping table. In this way, when it is necessary to process multiple message formats, each message format can be mapped to a suitable processing logic, so that chat messages of each message format can be processed without obstacles according to the mapped processing logic.

[0055] In this embodiment, an exemplary setting scheme of the mapping table is proposed: when the message format is text, the message content is the text message and text content sent; when the message format is picture, the message content is the picture sent and the picture size; when the message format is consent to archive the conversation content, the message content is the consent to archive the conversation content and the consent time sent; when the message format is disagreement to archive the conversation content, the message content is the disagreement to archive the conversation content and the disagreement time sent; when the message format is voice, the message content is the voice and voice content sent; when the message format is video, the message content is the video sent, the video playback length and the video file size; when the message format is business card, the business card company name and Business card identification; when the message format is location, the message content is the location information, location information, address information, longitude and latitude information and zoom ratio sent; when the message format is emoticon, the message content is the emoticon picture, emoticon picture size, emoticon picture height and emoticon picture width sent; when the message format is file, the message content is the file, file name and file size sent; when the message format is link, the message content is the link, link title and link description time sent; when the message format is mini program information, the message content is the mini program information, message title, mini program name and user name sent; when the message format is session record message, the message content is the session record message and session record title sent; when the message format is to-do When the message format is a voting message, the message content is that the to-do message and to-do content have been sent; when the message format is a voting message, the message content is that the voting message, voting subject and voting options have been sent; when the message format is a form filling message, the message content is that the form filling message, form filling name, form filling content and form filling item name have been sent; when the message format is a red envelope message, the message content is that the red envelope message, red envelope blessing, total number of red envelopes and total amount of red envelopes have been sent; when the message format is an online document message, the message content is that the online document message, document name and file creator have been sent; when the message format is a MarkDown format message, the message content is that the MarkDown format message and message content have been sent; when the message format is a graphic message, the message content is The picture and text message, picture and text title, picture and text description and picture and text content are sent; when the message format is a mixed message (including pictures and text), the message content is the mixed message sent, the number of pictures and texts contained in the mixed message and the text content; when the message format is an audio archive message, the message content is the audio archive message sent and the audio archive name; when the message format is an audio shared document message, the message content is the audio shared document message sent and the document sharing name; when the message format is an intercommunication red envelope message, the message content is the intercommunication red envelope message sent, the red envelope blessing, the total number of red envelopes and the total amount of red envelopes; when the message format is a video account message, the message content is the video account message sent, the video account account name and the video account message description;When the message format is audio and video call, the message content is the audio and video call and the call duration; when the message format is microdisk file, the message content is the microdisk file and the file name; when the message format is unknown message, the message content is the unknown message and the unknown message content. Here, the mapping table can be adaptively adjusted according to the actual application scenario, and this embodiment does not limit this. ;

[0056] Continue to refer Figure 1 Based on the mapping table proposed in this embodiment, the summary generation device can extract the message content corresponding to the required content dimension from any chat message according to the content dimension preset in the processing logic found for the chat message, wherein the length of the message content is less than the length of the chat message.

[0057] As mentioned above, since different message formats usually have different data structures, the data contents contained in different data structures are also different. Therefore, a corresponding content dimension can be set according to the data content contained in each message format, so as to process the chat message of each message format through the content dimension, and obtain the message content of the chat message after processing. Here, the setting method of the content dimension is not limited, and the extraction method used to extract the message content is not further limited. An optional implementation method of extracting the message content will be proposed later.

[0058] It can be understood that the preset content dimension in this embodiment does not refer to the entire content of the chat message, but rather selects part of the content in the chat message as the content dimension. Based on this, when the chat message is processed according to the content dimension, the redundant part of the summary generation in the chat message can be filtered out to obtain a message content whose length is shorter than the chat message, and then summary data can be generated according to the message content to improve the efficiency of summary generation.

[0059] Continue to refer Figure 1 The summary generation device can input the message contents generated for the multiple chat messages as input data into the target language model, and input preset prompt information into the target language model, wherein the prompt information includes task description information for instructing to generate a summary based on the input data. Here, the input method of inputting the message content into the target language model is not limited, and an optional implementation method will be provided later.

[0060] In this embodiment, the type of the target language model is not limited, and the type of the target language model has different adaptability according to the different implementation capabilities required by the target language model. No specific examples are given here for the type of the target language model.

[0061] As mentioned above, the prompt information in this embodiment is used to inform the target language model of the task to be performed and guide the target language model to output the corresponding content according to the task. The prompt information may include task description information. For example, when the target language model needs to generate a summary based on the conversation content, the prompt information may prompt the target language model that the current task to be performed is to generate a summary.

[0062] In addition, the prompt information can also prompt the target language model the usage scenario of the conversation content, the chat role that conducts the conversation in the usage scenario, and the conversation content before the conversation content, for reference by the target language model. Here, no further examples are given for the content included in the prompt information.

[0063] Continue to refer Figure 1 After the summary generation device inputs the input data and prompt information into the target language model, the target language model can perform natural language understanding on the input data according to the task description information contained in the prompt information to output processed data.

[0064] In this embodiment, the target language model can focus on the content related to the task description information in the input data. For example, the task description information is: Please extract the housekeeping service time booked by the customer from this input data. The task description information is used to indicate that the target language model needs to perform the task of extracting the housekeeping service time. The target language model can search for words and sentences related to time in the input data based on the task description information, thereby filtering out content in the input data that is not related to the task description information.

[0065] As mentioned above, part of the content in the chat message can be selected as the content dimension, so that when the chat message is processed according to the content dimension, the redundant part of the chat message for summary generation can be filtered out. In this embodiment, the task description information can be used to quickly filter out the content related to the task description information in the input data, to a certain extent avoiding the tedious process of manually checking the chat messages one by one, thereby improving the efficiency of information acquisition and further improving the efficiency of summary generation.

[0066] Continue to refer Figure 1 Based on the processed data output by the target language model, the target language model can send the processed data to the summary generating device, and the summary generating device can construct summary data corresponding to the chat record based on the processed data.

[0067] In this embodiment, there is no limitation on the construction method of the summary data corresponding to the chat record, and the summary data can be constructed according to the usage scenario of the chat record. An exemplary optional implementation method of constructing the summary data will be provided below.

[0068] It should be noted that, in this embodiment, the target language model does not sense whether the output processed data is summary data, but constructs summary data corresponding to the chat record through the summary generation device. In this way, the target language model can focus on processing the input data, avoiding the additional calculation amount caused by the target language model considering whether the output result is summary data, thereby improving the processing efficiency of the target language model. Moreover, since the processed data output by the target language model is not limited to summary data, this can make the processed data flexibly applied to a variety of scenarios, and it is also convenient to process the processed data in a variety of scenarios, thereby improving the versatility of the processed data.

[0069] In summary, through the above description, it can be known that in this embodiment, the chat record to be processed can be obtained, wherein the chat record contains multiple chat messages; determine the message format corresponding to each of the multiple chat messages; based on the mapping relationship between the preset message format and the processing logic, respectively search for the processing logic required for the multiple chat messages; for any chat message, according to the content dimension preset in the processing logic found for the chat message, extract the message content corresponding to the required content dimension from the chat message, wherein the length of the message content is less than the length of the chat message; input the message content generated for the multiple chat messages as input data to the target language model; input the preset prompt information to the target language model, wherein the prompt information contains task description information for indicating the generation of a summary based on the input data; use the target language model to perform natural language understanding on the input data according to the task description information to obtain processing data; based on the processing data, construct summary data corresponding to the chat record. Accordingly, based on the mapping relationship proposed in the embodiment of the present application, the chat message can be de-redundantly processed to compress the length of the chat message. In this way, when it is necessary to generate a summary for the chat record, a summary can be generated based on the compressed chat content, thereby saving the time of reading the entire chat message. Moreover, the summary of the chat record can be generated with the help of the target language model to realize the automation of summary generation, further saving the time of repeatedly reading the chat record, which can effectively improve the efficiency of summary generation.

[0070] In the above or below embodiments, a variety of implementations may be used to extract message content corresponding to the required content dimension from the chat message. An optional implementation is provided below.

[0071] In this optional implementation, it is proposed that: if the chat message is structured information, a target field corresponding to the content dimension is searched from the chat message; the target field and the field value under the target field are extracted as the message content.

[0072] Among them, structured information refers to information with a fixed format, which usually expresses the content according to a pre-set template or pattern. For example, in the scenario of housekeeping services, the structured information is {Message Type: Service Appointment; Message Sender: A; Message Receiver: B; Service Details: {Service Type: Home Cleaning; Appointment Time: C; Service Address: D}}. In the above example, the fields are Message Type, Message Sender, Message Receiver, Service Details, Service Type, Appointment Time, and Service Address; the field values ​​are Service Appointment, A, B, Home Cleaning, C, and D.

[0073] As mentioned above, the content dimension is to select part of the content in the chat message as the content dimension, that is, a specific content that is concerned according to specific needs. For example, if you want to know the details of the housekeeping service that the customer has booked, the content dimension is the service details; if you want to know the party who booked the housekeeping service, the content dimension is the message sender.

[0074] In this embodiment, based on the determination of the content dimension, the target field corresponding to the content dimension can be searched in the chat message. For example, if the content dimension is service details, the target field is each subfield under the service details, namely, service type, appointment time and service address. Based on this, the target field and the field value under the target field can be extracted as the message content. For example, the message content extracted from the above example is {Service type: home cleaning; appointment time: C; service address: D}. In actual operation, the above search and extraction operations can be completed by programming or with the help of data processing tools, which is not limited in this embodiment.

[0075] Accordingly, through the above-mentioned method of extracting message content, relevant information of the content dimension concerned in the chat message can be accurately presented to compress the length of the chat message, thereby improving the processing efficiency of the chat message and facilitating the generation of summary data using the extracted message content. Moreover, the chat message can be supported as a structured information type to ensure the flexibility of the process of extracting message content from the chat message.

[0076] In the above or following embodiments, since different message formats usually have different data structures, this embodiment adopts a method of classifying all message formats so as to centrally process the message formats under each category using a unified content dimension.

[0077] In this optional implementation method, it is proposed that: the message formats in the mapping relationship include a first group of message formats and a second group of message formats, each message format included in the first group of message formats depends on text transmission content, and each message format included in the second group of message formats depends on non-text transmission content; in the processing logic corresponding to each message format included in the first group of message formats, content class fields are used as preset content dimensions; in the processing logic corresponding to each message format included in the second group of message formats, attribute class fields are used as preset content dimensions.

[0078] Based on the aforementioned mapping table, the first group of message formats in this embodiment include text, voice, location, to-do messages, form filling messages, red envelope messages, MarkDown format messages, graphic and text messages, mixed messages, mutual red envelope messages, video number messages and unknown; the second group of message formats include pictures, consent to archive conversation content, disagreement to archive conversation content, video, business card, emoticon, file, link, mini program message, conversation record message, voting message, online document message, audio archive message, audio sharing document message, audio and video call and micro disk file. The content class field in this embodiment refers to the field related to the message content in the chat message, for example, the text content in the text format; the attribute class field refers to the field related to the attribute information in the chat message, for example, the picture size in the picture format.

[0079] Accordingly, by classifying the message formats, a unified processing logic can be used to extract and process chat messages in the same message format, thereby avoiding the need to write a processing program for each message format to simplify the processing flow. Moreover, the message content can be accurately extracted from the chat message based on the content dimension set for each type of message format, thereby reducing the problem of missing or erroneous information extraction caused by the diversity of message formats.

[0080] In the above or below embodiments, various implementations may be used to input the message contents generated for the plurality of chat messages as input data to the target language model. An optional implementation is provided below.

[0081] This optional implementation method proposes: splicing the message contents generated for multiple chat messages to obtain spliced ​​content; when the length of the spliced ​​content exceeds the length threshold required by the target language model, dividing the spliced ​​content into multiple content paragraphs, wherein the length of a single paragraph is less than the length threshold; and inputting the multiple content paragraphs into the target language model.

[0082] Since the target language model usually has a limit on the length of the input content, that is, a length threshold (for example, two thousand words), when the length of the input content exceeds the length threshold, it will cause the model to have too much calculation, reduce processing efficiency, and may even exceed the processing capacity of the model. Therefore, it is proposed in this embodiment that after splicing the contents of multiple messages, it is necessary to compare the length of the spliced ​​content with the length threshold required by the target language model. When the length of the spliced ​​content exceeds the length threshold, the spliced ​​content can be segmented to obtain multiple content paragraphs, so that the length of each content paragraph is less than the length threshold, so that each paragraph can meet the input requirements of the target language model. A segmentation method for segmenting the spliced ​​content will be provided later, and no further examples will be given here.

[0083] Accordingly, when the length of the spliced ​​content exceeds the length threshold, the spliced ​​content can be segmented to meet the input requirements of the target language model. Moreover, in the above-mentioned embodiment, the chat message can be de-redundant to compress the length of the chat message, and in this embodiment, in order to meet the input requirements of the target language model, the operation of segmenting the spliced ​​content can be understood as a secondary compression of the length of the chat message, which can save the processing time of the chat message and effectively improve the processing efficiency of the target language model.

[0084] As mentioned above, in this optional implementation, the spliced ​​content can be divided into multiple content paragraphs. In this regard, this optional implementation proposes a further implementation scheme: determining a target sentence that meets the length threshold in the spliced ​​content; using the end position of the target sentence as the segmentation position, and segmenting the spliced ​​content to obtain multiple content paragraphs.

[0085] In this embodiment, there is no limitation on the segmentation method for determining the segmentation position, and the segmentation may also be performed according to the chat topic in the chat message or according to the completeness of the sentence, and no further examples are given here.

[0086] In this embodiment, the target sentence can also be retained in the content paragraph after the target sentence to ensure that the two content paragraphs adjacent to the target sentence contain the target sentence, thereby achieving a certain transition and connection between the two adjacent content paragraphs in semantics, which facilitates the target language model to understand each content paragraph.

[0087] Based on this, the concatenated content can be segmented according to the above segmentation method to ensure that each content paragraph obtained after segmentation meets the length threshold requirement of the target language model, which is convenient for the subsequent input of the content paragraph into the target language model and the processing of the content paragraph by the target language model to ensure the semantic integrity on the basis of meeting the length limit.

[0088] In this further implementation scheme, it is proposed that: counting can be started from the first word of the spliced ​​content to determine the statistical position when the cumulative number of words reaches the length threshold; if the statistical position coincides with the end position of any content sentence included in the spliced ​​content, the content sentence is determined as the target sentence; if the statistical position is located in the middle of any content sentence included in the spliced ​​content, the content sentence before the content sentence is determined as the target sentence; statistics continue to be performed from the first word of the content sentence after the determined target sentence to continue searching for the target sentence, and the search is stopped after the remaining number of words in the spliced ​​content is less than the length threshold.

[0089] In this embodiment, the first word of the spliced ​​content can be used as the starting statistical point, and the number of statistical words can be calculated word by word starting from the starting statistical point. When the number of statistical words reaches the length threshold, the position where the length threshold is reached is determined as the statistical position.

[0090] If the statistical position is the position of a complete sentence, the complete sentence can be used as the target sentence, that is, the statistical position coincides with the end position of the content sentence, where the end position can be understood as the position of the last character (including punctuation marks) of the content sentence; if the statistical position is in the middle of the complete sentence, the sentence before the complete sentence can be used as the target sentence, thereby avoiding truncating a complete sentence in the middle, ensuring the semantic integrity of the target sentence, and facilitating the target language model to understand the target sentence.

[0091] In this embodiment, after the first target sentence is determined, the word count is continued in the same manner starting from the first word of the content sentence after the determined target sentence, and the next target sentence is searched again. The aforementioned processing is continuously repeated until the remaining number of words is less than the length threshold and the search for the target sentence is stopped, indicating that the parts of the spliced ​​content that can be reasonably divided have been processed.

[0092] In addition, this embodiment does not limit the implementation method of determining the target sentence. For example, the target sentence can also be determined according to the semantic integrity or by using punctuation marks. No further examples are given here.

[0093] Based on this, through such a method of cyclically searching for target sentences, the concatenated content can be reasonably segmented, so that the multiple content paragraphs finally obtained by segmentation can not only meet the input length requirements of the target language model, but also maintain semantic integrity, thereby improving the processing efficiency of the target language model for content paragraphs.

[0094] In the above or below embodiments, various implementations may be used to input the message contents generated for the plurality of chat messages as input data to the target language model. An optional implementation is provided below.

[0095] In this optional implementation, it is proposed that: the message contents generated for multiple chat messages are spliced ​​to obtain spliced ​​content; when the length of the spliced ​​content does not exceed the length threshold required by the target language model, the spliced ​​content is input into the target language model. Through the above processing method, multiple chat messages can be effectively spliced ​​to obtain spliced ​​content under the premise of complying with the input requirements of the target language model, so that the target language model can understand the semantics of the spliced ​​content without obstacles based on its own natural language processing capabilities.

[0096] In the above or following embodiments, based on the aforementioned content paragraphs, this optional implementation method proposes: if it is determined that the content has been divided into multiple content paragraphs after splicing, then based on the sequential relationship between the multiple content paragraphs, a previous adjacent content paragraph can be determined for any non-head content paragraph as the target content paragraph that needs to be dependent; in the task description information, it is indicated that when the target language model processes any non-head content paragraph, the processing data generated for the target content paragraph that needs to be dependent is used as reference information to generate processing data for the non-head content paragraph.

[0097] In this embodiment, it is proposed to determine a target content paragraph for any non-header content paragraph. For example, after splicing, the content is divided into three content paragraphs, which are marked as paragraph 1, paragraph 2 and paragraph 3. For paragraph 2, its previous adjacent content paragraph is paragraph 1, so paragraph 1 is the target content paragraph that paragraph 2 needs to depend on.

[0098] Based on this, in this embodiment, when processing any non-header content segment, the processed data generated by the target content segment may be used as reference information, so as to assist in generating the processed data of the non-header content segment using the reference information.

[0099] For example, suppose the target language model is tasked with analyzing the content paragraphs about housekeeping services in the chat record and generating processed data based on the content paragraphs. Assuming that paragraph 1 mainly describes the basic information of the customer's appointment for housekeeping services, such as the type of service being home cleaning and the appointment time, the target language model can process paragraph 1 and obtain the processed data of paragraph 1 after processing. Assuming that paragraph 2 mainly describes the customer's cleaning requirements for key cleaning areas, when the target language model processes paragraph 2, the task description information will instruct the target language model to refer to the previously generated processed data for paragraph 1, that is, the basic information for booking housekeeping services. In this way, the target language model can combine the previous service appointments to more comprehensively and accurately analyze the content of the cleaning requirements for key cleaning areas in paragraph 2, and can correlate the cleaning requirements for key cleaning areas with the type of service and appointment time of the appointment, thereby generating processing data for paragraph 2 that is more in line with the overall chat record context.

[0100] In addition, this embodiment can also set the model parameter temperature of the target language model to 0.1, so that the target language model can focus on the chat records, thereby generating concise and conventionally understood processed data, reducing irrelevant or inaccurate information caused by randomness in the processed data. For example, for a content paragraph about housekeeping services, a lower temperature value will enable the target language model to semantically understand the content paragraph according to the knowledge of the housekeeping service field, so as to output the processed data of the content paragraph. It can be understood that the temperature parameter is used to reduce the randomness of the model output results. The lower the temperature value, the worse the randomness of the model output results. This embodiment does not specifically limit the value of temperature.

[0101] Accordingly, through the above processing method, the target language model can fully consider the adjacent content paragraphs when processing each content paragraph, so that the target language model will not view each content paragraph in isolation, but can better understand the meaning of the current content paragraph based on the processing data of the previous content paragraph, avoiding information fragmentation, thereby ensuring the semantic coherence of the processing data output by the target language model. Moreover, using the processing data of the previous content paragraph as reference information can help the target language model more accurately analyze the information conveyed by the current content paragraph, thereby improving the accuracy of generating processing data for each content paragraph and the entire chat record.

[0102] In the above or below embodiments, various implementations may be used to construct summary data corresponding to the chat records based on the processed data. An optional implementation is provided below.

[0103] This optional implementation method proposes: if the length of the processed data is greater than the summary length threshold, the output data length in the task description information is adjusted; the processed data is used as input data and input into the target language model; the adjusted task description information is input into the target language model; the target language model is used to perform natural language understanding on the input data according to the adjusted task description information to obtain the adjusted processed data as the summary data corresponding to the chat record.

[0104] In this embodiment, a summary length threshold can be pre-set. The summary length threshold can be understood as a standard value used to describe whether the processed data meets the length requirement. The output data length in the task description information is used to describe the length range of the processed data expected to be output. If the length of the processed data output by the target language model is greater than the summary length threshold, it means that the generated processed data is too long. However, in the actual processing process, the length of the processed data will be affected by the content of the message. For example, when the message content is complex or there is a large amount of message content, although the output data length is set in the task description information, there will be a situation where the message content cannot be fully summarized within the range of the output data length, resulting in the length of the processed data being greater than the output data length.

[0105] Based on this, in this embodiment, in the case where the length of the processed data is greater than the length of the output data, the output data length is adjusted to expand the length range of the output processed data. For example, the output data length before adjustment is set within 20 words, and the output data length after adjustment can be set within 30 words.

[0106] For example, assuming that the summary length threshold is 100 words, if the length of the output processed data exceeds 100 words according to the task description information before adjustment (i.e., the output data length is set within 20 words), the output data length is adjusted (the output data length is set within 30 words), and the processed data and the adjusted task description information are input into the target language model, so as to generate the adjusted processed data according to the adjusted task description information using the target language model. It is worth noting that even if the length of the processed data does not exceed the summary length threshold, the processed data will be retained, that is, even if the processed data before adjustment exceeds the output data length (within 20 words), it will be retained. This embodiment does not limit the numerical values ​​of the summary length threshold and the output data length, and can be flexibly adjusted according to actual needs.

[0107] In this embodiment, if the target language model has an output abnormality problem, it can be retried. When the number of retries exceeds the retry threshold, the target language model can return a response carrying the reason for the output abnormality, so that adjustments can be made based on the reason for the output abnormality to facilitate the target language model to execute the next output process.

[0108] Based on this, in this embodiment, after the target language model outputs the processed data, the processed data can be returned to the summary generating device, and the summary generating device uses the processed data as summary data corresponding to the chat record.

[0109] Based on this, the above processing method can ensure that the processed data generated based on the input data can meet the expected requirements in length, so that the key information in the input data can be extracted to the greatest extent. The relationship between the expected length limit and the actual amount of text information is also considered, and the summary length threshold is further used to control the data length of the processed data. Moreover, the target language model does not perceive whether the output processed data is summary data, so that the target language model focuses on processing the input data, avoiding additional calculations, and thus improving the processing efficiency of the target language model.

[0110] Figure 3 A schematic diagram of a summary generation method provided by another exemplary embodiment of the present application, which can be applied to a terminal device or a server, is shown in FIG. Figure 3 , the method may include:

[0111] Step S301: obtaining a chat record to be processed, wherein the chat record includes a plurality of chat messages;

[0112] Step S302: Determine the message formats corresponding to the plurality of chat messages;

[0113] Step S303: based on the mapping relationship between the preset message format and the processing logic, searching for the processing logic required for the plurality of chat messages respectively;

[0114] Step S304: for any chat message, according to the content dimension preset in the processing logic found for the chat message, extract the message content corresponding to the required content dimension from the chat message, wherein the length of the message content is less than the length of the chat message;

[0115] Step S305: inputting the message contents generated for the plurality of chat messages respectively as input data into the target language model;

[0116] Step S306: inputting preset prompt information into the target language model, wherein the prompt information includes task description information for instructing to generate a summary based on the input data;

[0117] Step S307: using the target language model to perform natural language understanding on the input data according to the task description information to obtain processed data;

[0118] Step S308: Based on the processed data, construct summary data corresponding to the chat record.

[0119] In an optional embodiment, when extracting the message content corresponding to the required content dimension from the chat message, it includes:

[0120] If the chat message is structured information, searching the chat message for a target field corresponding to the content dimension;

[0121] The target field and the field value under the target field are extracted as the message content.

[0122] In an optional embodiment, the method also includes: the message formats in the mapping relationship include a first group of message formats and a second group of message formats, each message format included in the first group of message formats relies on text transmission content, and each message format included in the second group of message formats relies on non-text transmission content; in the processing logic corresponding to each message format included in the first group of message formats, a content class field is used as a preset content dimension; in the processing logic corresponding to each message format included in the second group of message formats, an attribute class field is used as a preset content dimension.

[0123] In an optional embodiment, when the message contents generated for the plurality of chat messages are respectively input as input data to the target language model, the method includes:

[0124] splicing the message contents generated for the plurality of chat messages respectively to obtain spliced ​​content;

[0125] When the length of the concatenated content exceeds the length threshold required by the target language model, the concatenated content is divided into a plurality of content paragraphs, wherein the length of a single paragraph is less than the length threshold;

[0126] The plurality of content paragraphs are input into the target language model.

[0127] In an optional embodiment, when the spliced ​​content is divided into a plurality of content sections, it includes:

[0128] Determining a target sentence that meets the length threshold in the concatenated content;

[0129] The end position of the target sentence is used as the segmentation position to segment the concatenated content to obtain the multiple content paragraphs.

[0130] In an optional embodiment, when determining the target sentence that meets the length threshold in the concatenated content, the method includes:

[0131] Starting from the first word of the spliced ​​content, determining the statistical position when the cumulative number of words reaches the length threshold;

[0132] If the statistical position coincides with the end position of any content sentence included in the spliced ​​content, the content sentence is determined as the target sentence;

[0133] If the statistical position is located in the middle of any content sentence included in the concatenated content, the content sentence before the content sentence is determined as the target sentence;

[0134] The target sentence is continuously searched by continuing to count the first word of the content sentence after the determined target sentence until the number of remaining words in the spliced ​​content is less than the length threshold and then stopping the search.

[0135] In an optional embodiment, the method further includes:

[0136] When the length of the concatenated content does not exceed the length threshold required by the target language model, the concatenated content is input into the target language model.

[0137] In an optional embodiment, the method further includes:

[0138] If it is determined that the spliced ​​content has been segmented into a plurality of content segments, then based on the order relationship between the plurality of content segments, a preceding adjacent content segment is determined for any non-head content segment as a target content segment to be dependent on;

[0139] The task description information indicates that when the target language model processes any non-header content paragraph, it uses the processing data generated for the target content paragraph that it depends on as reference information to generate processing data for the non-header content paragraph.

[0140] In an optional embodiment, when constructing summary data corresponding to the chat record based on the processed data, it includes:

[0141] If the length of the processed data is greater than the summary length threshold, adjusting the output data length in the task description information;

[0142] Inputting the processed data as input data into the target language model;

[0143] Inputting the adjusted task description information into the target language model;

[0144] The target language model is used to perform natural language understanding on the input data according to the adjusted task description information to obtain adjusted processed data as summary data corresponding to the chat record.

[0145] Figure 4 This is a schematic diagram of a structure of a summary generation device provided by another exemplary embodiment of the present application. The product form of the device can be a functional plug-in. Figure 4 , the device comprises:

[0146] The data acquisition unit 40 is used to acquire the chat record to be processed, wherein the chat record includes multiple chat messages;

[0147] A data determination unit 41, configured to determine a message format corresponding to each of the plurality of chat messages;

[0148] A logic search unit 42, configured to search for the processing logic required for the plurality of chat messages respectively based on a mapping relationship between a preset message format and a processing logic;

[0149] The logic processing unit 43 is used to extract, for any chat message, the message content corresponding to the required content dimension from the chat message according to the content dimension preset in the processing logic found for the chat message, wherein the length of the message content is less than the length of the chat message;

[0150] A first data input unit 44, configured to input the message contents generated for the plurality of chat messages respectively as input data into a target language model;

[0151] A second data input unit 45 is used to input preset prompt information into the target language model, wherein the prompt information includes task description information for instructing to generate a summary based on the input data;

[0152] A data processing unit 46, configured to perform natural language understanding on the input data according to the task description information using the target language model to obtain processed data;

[0153] The summary construction unit 47 is used to construct summary data corresponding to the chat record based on the processed data.

[0154] In an alternative embodiment, the device may be used to:

[0155] If the chat message is structured information, searching the chat message for a target field corresponding to the content dimension;

[0156] The target field and the field value under the target field are extracted as the message content.

[0157] In an optional embodiment, the device can be used for: the message formats in the mapping relationship include a first group of message formats and a second group of message formats, each message format included in the first group of message formats depends on text transmission content, and each message format included in the second group of message formats depends on non-text transmission content; in the processing logic corresponding to each message format included in the first group of message formats, a content class field is used as a preset content dimension; in the processing logic corresponding to each message format included in the second group of message formats, an attribute class field is used as a preset content dimension.

[0158] In an optional embodiment, the device may be used to: splice the message contents generated for the plurality of chat messages respectively to obtain spliced ​​content;

[0159] When the length of the concatenated content exceeds the length threshold required by the target language model, the concatenated content is divided into a plurality of content paragraphs, wherein the length of a single paragraph is less than the length threshold;

[0160] The plurality of content paragraphs are input into the target language model.

[0161] In an optional embodiment, the device may be used to: determine a target sentence that meets the length threshold in the concatenated content;

[0162] The end position of the target sentence is used as the segmentation position to segment the concatenated content to obtain the multiple content paragraphs.

[0163] In an optional embodiment, the device may be used to: start counting from the first word of the spliced ​​content to determine the statistical position when the cumulative word count reaches the length threshold;

[0164] If the statistical position coincides with the end position of any content sentence included in the spliced ​​content, the content sentence is determined as the target sentence;

[0165] If the statistical position is located in the middle of any content sentence included in the concatenated content, the content sentence before the content sentence is determined as the target sentence;

[0166] The target sentence is continuously searched by continuing to count the first word of the content sentence after the determined target sentence until the number of remaining words in the spliced ​​content is less than the length threshold and then stopping the search.

[0167] In an optional embodiment, the device may be used to: when the length of the concatenated content does not exceed a length threshold required by the target language model, input the concatenated content into the target language model.

[0168] In an optional embodiment, the device may be used to: if it is determined that the spliced ​​content has been segmented into a plurality of content segments, then based on the sequence relationship between the plurality of content segments, determine a preceding adjacent content segment for any non-header content segment as a target content segment to be dependent on;

[0169] The task description information indicates that when the target language model processes any non-header content paragraph, it uses the processing data generated for the target content paragraph that it depends on as reference information to generate processing data for the non-header content paragraph.

[0170] In an optional embodiment, the apparatus may be used to: if the length of the processed data is greater than a summary length threshold, adjust the output data length in the task description information;

[0171] Inputting the processed data as input data into the target language model;

[0172] Inputting the adjusted task description information into the target language model;

[0173] The target language model is used to perform natural language understanding on the input data according to the adjusted task description information to obtain adjusted processed data as summary data corresponding to the chat record.

[0174] It is worth noting that the technical details in the above-mentioned embodiments of the summary generation device can refer to the relevant description of the summary generation method in the above-mentioned embodiments. In order to save space, they will not be repeated here, but this should not cause loss of the protection scope of this application.

[0175] Figure 5 A schematic diagram of a computing platform provided as another exemplary embodiment of the present application. Figure 5 As shown, the computing platform in practice includes: a memory 50 and a processor 51.

[0176] The memory 50 is used to store computer programs and can be configured to store various other data to support operations on the computing platform. Examples of such data include instructions for any application or method operating on the computing platform, contact data, phone book data, messages, pictures, videos, etc.

[0177] The processor 51 is coupled to the memory 50 and is used to execute the computer program in the memory 50 to implement the aforementioned summary generation method.

[0178] Further, if Figure 5 As shown, the computing platform also includes: a communication component 52, a display 53, a power component 54, an audio component 55 and other components. Figure 5 Only some components are shown schematically, which does not mean that the computing platform only includes Figure 5 In addition, Figure 5 The components in the dashed box are optional components, not mandatory components, and the specific components depend on the product form of the working node. The working node of this embodiment can be implemented as a terminal device such as a desktop computer, a laptop computer, a smart phone, or an IOT device, or a server-side device such as a conventional server, a cloud server, or a server array. If the working node of this embodiment is implemented as a terminal device such as a desktop computer, a laptop computer, a smart phone, etc., it can include Figure 5If the working node of this embodiment is implemented as a server device such as a conventional server, a cloud server or a server array, it may not include Figure 5 Components within the dashed box.

[0179] The above-mentioned memory 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 storage, flash memory, magnetic disk or optical disk.

[0180] The communication component is configured to facilitate wired or wireless communication between the device where the communication component is located and other devices. The device where the communication component is located can access a wireless network based on a communication standard, such as a mobile communication network such as 2G, 3G, 4G / LTE, 5G, or a combination thereof. In an exemplary embodiment, the communication component receives a broadcast signal or broadcast-related information from an external broadcast management system via a broadcast channel.

[0181] The above-mentioned display includes a screen, and the screen may include a liquid crystal display (LCD) and a touch panel (TP). If the screen includes a touch panel, the screen may be implemented as a touch screen to receive input signals from a user. The touch panel includes one or more touch sensors to sense touch, slide, and gestures on the touch panel. The touch sensor may not only sense the boundary of a touch or slide action, but also detect the duration and pressure associated with the touch or slide operation.

[0182] The power supply assembly provides power to various components of the device where the power supply assembly is located. The power supply assembly may include a power management system, one or more power supplies, and other components associated with generating, managing, and distributing power to the device where the power supply assembly is located.

[0183] The above-mentioned audio component can be configured to output and / or input audio signals. For example, the audio component includes a microphone (Microphone, MIC), and when the device where the audio component is located is in an operating mode, such as a call mode, a recording mode, and a speech recognition mode, the microphone is configured to receive an external audio signal. The received audio signal can be further stored in a memory or sent via a communication component. In some embodiments, the audio component also includes a speaker for outputting an audio signal.

[0184] Accordingly, the embodiment of the present application also provides a computer-readable storage medium storing a computer program, and when the computer program is executed by a processor, the processor is enabled to implement each step in the above method embodiment. Among them, the computer-readable storage medium includes volatile or non-volatile or a combination thereof, and can be removable or non-removable. Examples of computer-readable storage media include, but are not limited to, phase-change random access memory (Phase-change Random Access Memory, PRAM), static random access memory (SRAM), dynamic random access memory (Dynamic Random Access Memory, DRAM), other types of random access memory (Random-Access Memory, RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), flash memory or other memory technology, read-only compact disc read-only memory (CD-ROM), digital versatile disc (Digital Video Disc, DVD) or other optical storage, magnetic cassette, tape disk storage or other magnetic storage device or any other non-transmission medium.

[0185] It should also be noted that the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, commodity or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, commodity or device. In the absence of more restrictions, the elements defined by the sentence "comprises a ..." do not exclude the existence of other identical elements in the process, method, commodity or device including the elements.

[0186] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with the relevant laws, regulations and standards of the relevant countries and regions, and provide corresponding operation entrances for users to choose to authorize or refuse.

[0187] The above is only the embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the present application may have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included in the protection scope of the present application.

Claims

1. A method for generating a summary, characterized in that: include: Obtaining a chat record to be processed, wherein the chat record includes multiple chat messages; Determining message formats corresponding to each of the plurality of chat messages; Based on the mapping relationship between the preset message format and the processing logic, searching for the processing logic required for the multiple chat messages respectively; For any chat message, according to the content dimension preset in the processing logic found for the chat message, extract the message content corresponding to the required content dimension from the chat message, wherein the length of the message content is less than the length of the chat message; Inputting the message contents generated for the plurality of chat messages respectively as input data into the target language model; Inputting preset prompt information into the target language model, wherein the prompt information includes task description information for instructing to generate a summary based on the input data; Using the target language model to perform natural language understanding on the input data according to the task description information to obtain processed data; Based on the processed data, summary data corresponding to the chat record is constructed.

2. The method according to claim 1, characterized in that Extracting message content corresponding to the required content dimension from the chat message includes: If the chat message is structured information, searching the chat message for a target field corresponding to the content dimension; The target field and the field value under the target field are extracted as the message content.

3. The method according to claim 1 or 2, characterized in that: The message formats in the mapping relationship include a first group of message formats and a second group of message formats, each message format included in the first group of message formats depends on textual transmission content, and each message format included in the second group of message formats depends on non-textual transmission content; In the processing logic corresponding to each message format included in the first group of message formats, the content class field is used as the preset content dimension; in the processing logic corresponding to each message format included in the second group of message formats, the attribute class field is used as the preset content dimension.

4. The method according to claim 1, characterized in that Inputting the message contents generated for the plurality of chat messages respectively as input data into the target language model includes: splicing the message contents generated for the plurality of chat messages respectively to obtain spliced ​​content; When the length of the concatenated content exceeds the length threshold required by the target language model, the concatenated content is divided into a plurality of content paragraphs, wherein the length of a single paragraph is less than the length threshold; The plurality of content paragraphs are input into the target language model.

5. The method according to claim 4, characterized in that The spliced ​​content is divided into multiple content sections, including: Determining a target sentence that meets the length threshold in the concatenated content; The end position of the target sentence is used as the segmentation position to segment the concatenated content to obtain the multiple content paragraphs.

6. The method according to claim 5, characterized in that Determining a target sentence that meets the length threshold in the concatenated content includes: Starting from the first word of the spliced ​​content, determining the statistical position when the cumulative number of words reaches the length threshold; If the statistical position coincides with the end position of any content sentence included in the spliced ​​content, the content sentence is determined as the target sentence; If the statistical position is located in the middle of any content sentence included in the concatenated content, the content sentence before the content sentence is determined as the target sentence; The target sentence is continuously searched by continuing to count the first word of the content sentence after the determined target sentence until the number of remaining words in the spliced ​​content is less than the length threshold and then stopping the search.

7. The method according to claim 4, characterized in that The method further comprises: When the length of the concatenated content does not exceed the length threshold required by the target language model, the concatenated content is input into the target language model.

8. The method according to claim 4, characterized in that The method further comprises: If it is determined that the spliced ​​content has been segmented into a plurality of content segments, then based on the order relationship between the plurality of content segments, a preceding adjacent content segment is determined for any non-head content segment as a target content segment to be dependent on; The task description information indicates that when the target language model processes any non-header content paragraph, it uses the processing data generated for the target content paragraph that it depends on as reference information to generate processing data for the non-header content paragraph.

9. The method according to claim 1, characterized in that: Based on the processed data, construct summary data corresponding to the chat record, including: If the length of the processed data is greater than the summary length threshold, adjusting the output data length in the task description information; Inputting the processed data as input data into the target language model; Inputting the adjusted task description information into the target language model; The target language model is used to perform natural language understanding on the input data according to the adjusted task description information to obtain adjusted processed data as summary data corresponding to the chat record.

10. A summary generation device, characterized in that: include: A data acquisition unit, used to acquire a chat record to be processed, wherein the chat record includes a plurality of chat messages; A data determination unit, used to determine the message formats corresponding to the plurality of chat messages; A logic search unit, used to search for the processing logic required for the plurality of chat messages respectively based on a mapping relationship between a preset message format and a processing logic; a logic processing unit, configured to extract, for any chat message, message content corresponding to a required content dimension from the chat message according to a content dimension preset in a processing logic found for the chat message, wherein the length of the message content is less than the length of the chat message; A first data input unit, configured to input the message contents generated for the plurality of chat messages respectively as input data into a target language model; A second data input unit, configured to input preset prompt information into the target language model, wherein the prompt information includes task description information for instructing to generate a summary based on the input data; A data processing unit, configured to perform natural language understanding on the input data according to the task description information using the target language model to obtain processed data; A summary construction unit is used to construct summary data corresponding to the chat record based on the processed data.

11. A computing platform, characterized in that: include: Memory and processor; The memory is used to store one or more computer instructions; The processor is coupled to the memory and configured to execute the one or more computer instructions to perform the abstract generation method according to any one of claims 1 to 9.

12. A computer-readable storage medium storing computer instructions, characterized in that: When the computer instructions are executed by one or more processors, the one or more processors are caused to execute the abstract generation method according to any one of claims 1 to 9.

Images