User text summarization method and device and electronic equipment

Abstract

The invention provides a user text summarization method and device and electronic equipment, and the method comprises the steps: carrying out the clustering of a to-be-processed user text, and obtaining a plurality of class clusters; aiming at each class cluster in the plurality of class clusters, performing question abstract on all user texts under the class cluster by using a large language model to obtain a question abstract text; predicting the matching probability between all user texts and question abstract texts under the cluster by using a small network model, wherein the model parameters of the small network model are less than the model parameters of the large language model; and determining a final abstract result according to the matching probability and the question abstract text. In the implementation process of the scheme, the balance between the accuracy and the real-time performance of the text abstract of the user is considered by combining the large language model with strong generation capability and theoretical force and the small network model with the real-time performance.

Description

Technical Field

[0001] The present application relates to the technical fields of deep learning, neural networks and text processing, and in particular to a user text summarization method, device and electronic device. Background Art

[0002] At present, the accuracy of the open source Large Language Model as a general model can reach about 70%, but when faced with text summarization tasks in specific professional fields, they often perform unsatisfactorily. This is mainly because the data that the large language model is exposed to during the pre-training stage is too generalized and lacks a deep understanding of a specific field. Therefore, when you want to use a large language model to summarize the text input by users in a professional field, you still need to fine-tune the large language model again to meet the standards for use in commercial scenarios. However, this process also brings new challenges, especially the balance between the accuracy and real-time performance of text summarization. Improving accuracy usually means more training time and computing resources, and the need for real-time performance conflicts with the complexity and training time of the model. Summary of the invention

[0003] The purpose of the embodiments of the present application is to provide a user text summarization method, device and electronic device, which are used to improve the problem of balancing the accuracy and real-time performance of text summarization.

[0004] The embodiment of the present application provides a user text summarization method, including: clustering the user text to be processed to obtain multiple clusters; for each cluster in the multiple clusters, using a large language model to perform question summarization on all user texts under the cluster to obtain question summary text; using a small network model to predict the matching probability between all user texts under the cluster and the question summary text, the model parameters of the small network model are less than the model parameters of the large language model; determining the final summary result according to the matching probability and the question summary text. In the implementation process of the above scheme, by combining a large language model with powerful generation and comprehension capabilities and a small network model with real-time performance, a different approach from training a large language model alone is provided. When the large language model generates high-quality and accurate question summary text based on the user text, the small network model continues to be used as a supervisor to predict the matching probability between the user text under the cluster and the question summary text, and reduces irrelevant or redundant information according to the matching probability, thereby effectively taking into account the balance between the accuracy and real-time performance of the user text summary.

[0005] Optionally, in an embodiment of the present application, before clustering the user text to be processed, it also includes: if the order number is identified from the return order data, questionnaire survey data, news public opinion data, user evaluation data, online consultation data or customer service work order data, then the user text is extracted; if the product identification is identified from the user evaluation data, news public opinion data or online consultation data, then the user text is extracted from the user evaluation data, news public opinion data or online consultation data. In the implementation process of the above scheme, data from different sources (such as return order data, user evaluation data, etc.) are associated through order numbers or product identifications, which can more comprehensively analyze user behavior and product performance, thereby drawing more accurate conclusions.

[0006] Optionally, in an embodiment of the present application, clustering user texts includes: segmenting user texts to obtain multiple words; vectorizing and encoding multiple words to obtain word encoding vectors; vectorizing and encoding user texts to obtain text encoding vectors, and splicing the text encoding vectors with the word encoding vectors to obtain spliced ​​encoding vectors; calculating feature similarities between spliced ​​encoding vectors, and clustering user texts based on feature similarities. In the implementation of the above scheme, by converting words into vectors, the semantic relationship between words can be captured, and converting the entire text into a vector can capture the overall meaning of the text, thereby further splicing the word encoding vectors with the text encoding vectors to obtain a richer text representation, which can better retain the details at the word level, while also retaining the overall features at the text level, thereby improving the accuracy of clustering.

[0007] Optionally, in an embodiment of the present application, after obtaining multiple words, it also includes: for each word in the multiple words, obtaining the frequency of the word in the user text, and / or counting the total number of user texts corresponding to the word; if the frequency of the word in the user text is greater than a preset frequency threshold, and / or the total number of user texts corresponding to the word is greater than a preset number threshold, then an early warning is issued for the word. In the implementation process of the above scheme, by counting the frequency of words and the number of texts, user feedback can be analyzed more finely to discover potential trends and patterns. By setting preset frequency thresholds and quantity thresholds, important words can be automatically screened out to reduce the workload of manual review. The early warning mechanism can realize real-time monitoring, helping enterprises to discover problems and respond in the first place. Enterprises can notice potential risks before the problem expands, and take preventive measures, thereby improving the accuracy of data analysis and data processing efficiency.

[0008] Optionally, in an embodiment of the present application, a large language model is used to summarize the questions of all user texts under the cluster, including: for each of the multiple clusters, all user texts under the cluster are labeled according to preset category labels to obtain multiple text labels; for each text label in the multiple text labels, the large language model is used to summarize the questions of the user text corresponding to the text label. In the implementation process of the above scheme, a large amount of user text can be automatically classified into different clusters through automatic labeling, which reduces the workload of manual classification, and automatic labeling can significantly improve the speed of text classification; further, using a large language model for problem summarization can automatically extract key information from user texts without manual reading and summarizing one by one, thereby greatly improving the efficiency of text processing through automatic labeling and automatic summarization.

[0009] Optionally, in an embodiment of the present application, a preset category label includes: a product category label and a question category label; labeling all user texts under the cluster according to the preset category label includes: structuring the user text to obtain structured content; judging whether the structured content meets preset conditions, the preset conditions include: the similarity between the product category in the structured content and the product category label is greater than a first preset similarity threshold, and the similarity between the question category in the structured content and the question category label is greater than a second preset similarity threshold; if so, determining the question category label as a text label. In the implementation process of the above scheme, free text can be converted into a structured data format through structured processing, making it easier for computers to understand key information in the text, and by judging the similarity between the product category in the structured content and the product category label, the product-related information in the text can be more accurately understood.

[0010] Optionally, in an embodiment of the present application, after obtaining multiple text tags, it also includes: for each text tag in the multiple text tags, obtaining the total number of user texts corresponding to the text tag; if the total number of user texts corresponding to the text tag is greater than a preset number threshold, then issuing an early warning for the text tag. In the implementation of the above scheme, through the statistics of text tags, user feedback can be analyzed more finely to discover potential trends and patterns. Through the early warning mechanism, enterprises can notice potential risks before the problem expands, thereby taking preventive measures to help enterprises monitor market public opinion, promptly discover negative information and respond.

[0011] Optionally, in an embodiment of the present application, before using the small network model to predict the matching probability between all user texts under the cluster and the question summary text, it also includes: obtaining multiple sample texts, summary labels and matching labels for each sample text; using the small network model to predict the matching probability between each sample text and the summary label of the sample text to obtain the predicted probability; calculating the loss value between the predicted probability and the matching label; and training the small network model according to the loss value. In the implementation process of the above scheme, by predicting the matching probability between the sample text and the summary label, the model can learn the intrinsic connection between the sample text and the summary label, and calculating the loss value between the predicted probability and the matching label can prompt the model to continuously adjust the weights to better fit the training data, thereby improving the generalization ability on unseen data.

[0012] Optionally, in an embodiment of the present application, the final summary result is determined based on the matching probability and the question summary text, including: judging whether the matching probability is greater than a preset probability threshold; if so, determining the question summary text as the final summary result, otherwise, using a large language model to re-summarize all user texts under the cluster, obtaining a re-summarized text, and determining the re-summarized text as the final summary result. In the implementation process of the above scheme, by setting a probability threshold, it can be ensured that the summary is only adopted when the model is very confident about the matching probability, which can significantly improve the quality and accuracy of the final summary. For those summaries with low matching probabilities, re-summarizing using a large language model can further optimize the summary content and improve its relevance and clarity of expression.

[0013] Optionally, in the embodiment of the present application, after obtaining multiple clusters, it also includes: for each cluster in the multiple clusters, if the total number of user texts under the cluster is greater than a preset number threshold, an early warning is issued for the problem summary text of the cluster. In the implementation process of the above scheme, by setting a number threshold to trigger the early warning mechanism, a sudden increase in the attention of a topic or issue can be discovered in time, which helps to identify potential important trends or events at an early stage, thereby improving the ability and accuracy of anomaly detection.

[0014] The embodiment of the present application also provides a user text summarization device, including: a user text clustering module, used to cluster the user text to be processed to obtain multiple clusters; a text question summarization module, used to use a large language model to perform question summarization on all user texts under each cluster in the multiple clusters to obtain a question summary text; a matching probability prediction module, used to use a small network model to predict the matching probability between all user texts under the cluster and the question summary text, the weight parameter of the small network model is less than the weight parameter of the large language model; a summary result determination module, used to determine the final summary result according to the matching probability and the question summary text.

[0015] Optionally, in an embodiment of the present application, the user text summary device also includes: a first text extraction submodule, which is used to extract user text from return order data, questionnaire survey data, news public opinion data, user evaluation data, online consultation data or customer service work order data if an order number is identified from return order data, questionnaire survey data, news public opinion data, user evaluation data, online consultation data or customer service work order data; a second text extraction submodule, which is used to extract user text from user evaluation data, news public opinion data or online consultation data if a product logo is identified from user evaluation data, news public opinion data or online consultation data.

[0016] Optionally, in an embodiment of the present application, a user text clustering module includes: a user text segmentation submodule, which is used to segment user text to obtain multiple words; a word encoding vector acquisition submodule, which is used to vectorize and encode multiple words to obtain word encoding vectors; a word encoding vector splicing submodule, which is used to vectorize and encode user text to obtain a text encoding vector, and splice the text encoding vector with the word encoding vector to obtain a spliced ​​encoding vector; a feature similarity clustering submodule, which is used to calculate the feature similarity between the spliced ​​encoding vectors and cluster the user text according to the feature similarity.

[0017] Optionally, in an embodiment of the present application, the user text clustering module also includes: a word frequency text statistics submodule, which is used to obtain the word frequency of each word in the user text for each of the multiple words, and / or to count the total number of user texts corresponding to the word; a word warning submodule, which is used to issue a warning for the word if the word frequency of the word in the user text is greater than a preset frequency threshold, and / or the total number of user texts corresponding to the word is greater than a preset number threshold.

[0018] Optionally, in an embodiment of the present application, the text question summary module includes: a text label acquisition submodule, which is used to label all user texts under each cluster among multiple clusters according to preset category labels to obtain multiple text labels; a label text summary submodule, which is used to use a large language model to perform question summary on the user text corresponding to each text label among the multiple text labels.

[0019] Optionally, in an embodiment of the present application, the preset category labels include: product category labels and question category labels; the text label acquisition submodule includes: a structured content acquisition unit, which is used to structure the user text and obtain structured content; a structured content judgment unit, which is used to judge whether the structured content meets preset conditions, and the preset conditions include: the similarity between the product category in the structured content and the product category label is greater than a first preset similarity threshold, and the similarity between the question category in the structured content and the question category label is greater than a second preset similarity threshold; a text label determination unit, which is used to determine the question category label as a text label if the structured content meets the preset conditions.

[0020] Optionally, in an embodiment of the present application, the text problem summary module also includes: a text quantity acquisition submodule, which is used to obtain, for each text tag among multiple text tags, the total number of user texts corresponding to the text tag; and a text tag warning submodule, which is used to issue a warning for the text tag if the total number of user texts corresponding to the text tag is greater than a preset quantity threshold.

[0021] Optionally, in an embodiment of the present application, the user text summarization device also includes: a text label acquisition module, used to obtain multiple sample texts, a summary label and a matching label for each sample text; a prediction probability acquisition module, used to use a small network model to predict the matching probability between the summary label of the sample text and each sample text, and obtain the prediction probability; a loss value calculation module, used to calculate the loss value between the prediction probability and the matching label; and a network model training module, used to train the small network model according to the loss value.

[0022] Optionally, in an embodiment of the present application, the summary result determination module includes: a matching probability judgment submodule, used to judge whether the matching probability is greater than a preset probability threshold; a summary result determination submodule, used to determine the question summary text as the final summary result if the matching probability is greater than the preset probability threshold, and if the matching probability is less than or equal to the preset probability threshold, use the large language model to re-summarize all user texts under the cluster to obtain a re-summary text, and determine the re-summary text as the final summary result.

[0023] Optionally, in an embodiment of the present application, the user text summary device further includes: a problem summary warning module, which is used to issue a warning for the problem summary text of each cluster among the multiple clusters if the total number of user texts under the cluster is greater than a preset number threshold.

[0024] An embodiment of the present application further provides an electronic device, including: a processor and a memory, wherein the memory stores machine-readable instructions executable by the processor, and the machine-readable instructions execute the method described above when executed by the processor.

[0025] An embodiment of the present application further provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the method described above is executed.

[0026] The embodiment of the present application also provides a computer program product, including: a computer program or a computer instruction, and the computer program or the computer instruction executes the method described above when executed by a processor. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the embodiments of the present application will be briefly introduced below. It should be understood that the following drawings only show certain embodiments of the embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without creative work.

[0028] Figure 1 A flowchart of a user text summarization method provided in an embodiment of the present application is shown;

[0029] Figure 2 A schematic diagram of a use case for managing relationships using early warning rules provided in an embodiment of the present application is shown;

[0030] Figure 3 A schematic diagram of the process flow of a user text summarization device provided in an embodiment of the present application is shown;

[0031] Figure 4 A schematic diagram of the structure of an electronic device provided in an embodiment of the present application is shown. DETAILED DESCRIPTION

[0032] To make the purpose, technical scheme and advantages of the embodiment of the present application clearer, the technical scheme in the embodiment of the present application will be clearly and completely described below in conjunction with the drawings in the embodiment of the present application. It should be understood that the drawings in the embodiment of the present application only serve the purpose of explanation and description, and are not used to limit the protection scope of the embodiment of the present application. In addition, it should be understood that the schematic drawings are not drawn in real proportion. The flowchart used in the embodiment of the present application shows the operation implemented according to some embodiments of the embodiment of the present application. It should be understood that the operation of the flowchart can be implemented out of order, and the steps without logical context can be reversed in order or implemented simultaneously. In addition, those skilled in the art can add one or more other operations to the flowchart under the guidance of the content of the embodiment of the present application, or remove one or more operations from the flowchart.

[0033] In addition, the described embodiments are only a part of the embodiments of the present application, rather than all embodiments. The components of the embodiments of the present application described and shown in the drawings generally here can be arranged and designed in various configurations. Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the embodiments of the present application claimed, but merely represents the selected embodiments of the embodiments of the present application.

[0034] It is understandable that the "first" and "second" in the embodiments of the present application are used to distinguish similar objects. Those skilled in the art will understand that the words "first" and "second" do not limit the quantity and execution order, and the words "first" and "second" do not necessarily limit the difference. In the description of the embodiments of the present application, the term "and / or" is merely a description of the association relationship of associated objects, indicating that there may be three relationships, such as A and / or B, which can represent: A exists alone, A and B exist at the same time, and B exists alone. In addition, the character " / " in this article generally indicates that the related objects before and after are in an "or" relationship. The term "multiple" refers to more than two (including two), and similarly, "multiple groups" refers to more than two groups (including two groups).

[0035] It should be noted that the user text summarization method provided in the embodiment of the present application can be executed by an electronic device, where the electronic device refers to a device terminal having the function of executing a computer program or the above-mentioned server, and the device terminal is, for example, a smart phone, a personal computer, a tablet computer, a personal digital assistant, or a mobile Internet device, etc. A server refers to a device that provides computing services through a network, and the server is, for example, an x86 server and a non-x86 server, and the non-x86 server includes: a mainframe, a minicomputer, and a UNIX server.

[0036] The following is an example of an application scenario in which the user text summary method is applicable: when it is necessary to analyze and process massive amounts of user voice data, the massive amounts of user voice data can be identified first to obtain a large amount of user text, and the user text summary method can be used to process a large amount of user text to generate high-quality and accurate problem summary texts, so as to strike a balance between the accuracy and real-time nature of the user text summary. Clustering and other analyses of user texts can help understand customers' perceptions and interactions with brands, products, and services throughout the life cycle of the service, thereby discovering and solving customer problems and providing customers with a better experience.

[0037] See also Figure 1 The flowchart of the user text summarization method provided by the embodiment of the present application is shown; the embodiment of the present application provides a user text summarization method, including:

[0038] Step S110: clustering the user text to be processed to obtain multiple clusters.

[0039] User text refers to the voice content text emitted by users or customers. The voice content emitted by customers is also called Voice of the Customer (VOC), such as feedback on products and services, including comments, expectations, preferences, etc. If the data form of VOC is sound data, then VOC can be first recognized as user text, for example, using whisper to recognize user text from sound data.

[0040] Step S120: For each cluster among the multiple clusters, use the large language model to perform question summary on all user texts under the cluster to obtain question summary text.

[0041] It is understandable that since the large language model has stronger semantic understanding and generation capabilities and its model parameters are more than those of the small network model, the effect of using a large language model for question summarization is better and the accuracy of question summaries is higher.

[0042] Step S130: Use the small network model to predict the matching probability between all user texts and question summary texts under the cluster, and the model parameters of the small network model are less than the model parameters of the large language model.

[0043] It is understandable that since the number of model parameters of the small network model is less than the number of model parameters of the large language model, when the small network model performs predictive analysis on the matching probability between user text and question summary text, compared with the predictive analysis performed by the large language model, the small network model consumes less resources and takes less time. Therefore, the trained small network model can be used in scenarios with higher real-time requirements.

[0044] Step S140: Determine the final summary result according to the matching probability and the question summary text.

[0045] Since the large language model has a higher accuracy in question summarization and the small network model has better real-time performance, the final summary result is determined by combining the question summary text generated by the large language model and the matching probability predicted by the small network model, which can better strike a balance between the accuracy and real-time performance of the text summary.

[0046] In the process of implementing the above solution, by combining a large language model with strong generation and comprehension capabilities and a small network model with real-time performance, a different approach from training a large language model alone is provided. While the large language model generates high-quality and accurate question summary text based on user text, the small network model continues to act as a supervisor to predict the matching probability between user text and question summary text under the cluster, and reduces irrelevant or redundant information based on the matching probability, thereby effectively striking a balance between the accuracy and real-time performance of the user text summary.

[0047] As an optional implementation of the above-mentioned user text summarization method, before clustering the user text to be processed, user text can also be extracted from multiple channel data, such as return order data, questionnaire survey data, news public opinion data (including institutional public opinion data and peripheral public opinion data), user evaluation data, online consultation data (such as online service record text or recording, etc.) or customer service work order data, etc. In addition, user text can also be extracted from channel data such as product evaluation data and Net Promoter Score (NPS) feedback data.

[0048] Optionally, identifying user text associated with the order number according to the order number, this implementation may include:

[0049] Step S101: If an order number is identified from return order data, questionnaire survey data, news and public opinion data, user evaluation data, online consultation data and / or customer service work order data, user text is extracted.

[0050] In the implementation of the above-mentioned step S101, by monitoring the order number in the return order data, potential quality problems can be discovered in time, and measures can be taken in advance to avoid greater losses. Sentiment analysis can also be performed on user evaluation data identified by product identification to help companies understand users' emotional tendencies towards specific products and guide product improvement directions.

[0051] Optionally, the user text associated with the commodity identification is identified according to the commodity identification. This implementation may include:

[0052] Step S102: If the product logo is identified from the user evaluation data, news public opinion data or online consultation data, then the user text is extracted from the user evaluation data, news public opinion data or online consultation data.

[0053] In the implementation of the above-mentioned steps S101 and S102, by identifying the order number or the product identification, the enterprise can provide more personalized customer service, because the enterprise can quickly identify the problem based on the order number or the product identification, that is, it can quickly locate the specific transaction record, so that it is easier to find the problem, and the enterprise can respond to user feedback more quickly and improve the efficiency of problem solving. Furthermore, by monitoring the product identification in the news public opinion data, negative reports can be discovered in time, and measures can be taken to conduct crisis public relations and reduce negative impacts. In the implementation process of the above-mentioned scheme, data from different sources (such as return order data, user evaluation data, etc.) are associated through order numbers or product identifications, so that user behavior and product performance can be analyzed more comprehensively, thereby drawing more accurate conclusions.

[0054] As an optional implementation of the above step S110, when clustering the user texts, clustering can be performed according to the similarity of the encoding features. This implementation may include:

[0055] Step S111: segment the user text to obtain multiple words.

[0056] The implementation method of the above step S111 is, for example, to use Jieba word segmentation tool library, NLTK, Spacy or Gensim and other Python tools to segment the user text into words, also known as tokenization, which is the process of dividing a continuous text sequence into independent vocabulary units, that is, breaking the user text from a long sentence into separate words or phrases to obtain multiple words. Optionally, after obtaining multiple words, the multiple words can also be cleaned, for example, by removing stop words or identifying the frequency of hot words, etc.

[0057] Step S112: vectorize and encode multiple words to obtain word encoding vectors.

[0058] It is understandable that converting words or text into vector form can speed up subsequent processing, especially on large-scale data sets. This is because vector operations are usually more efficient than direct comparison of text strings, and vectorized representation allows the use of parallel computing technology to speed up the processing process, especially in a distributed computing environment, thereby improving the processing efficiency of clustering analysis.

[0059] Step S113: vectorize and encode the user text to obtain a text encoding vector, and concatenate the text encoding vector with the word encoding vector to obtain a concatenated encoding vector.

[0060] The implementation method of the above steps S112 to S113 is, for example: using the Bag of Words (BoW), Word2Vec, GloVe and word embeddings to vectorize and encode multiple words to obtain word encoding vectors, and using document embeddings such as Doc2Vec or encoders such as BERT, RoBERTa and GPT to vectorize and encode user text to obtain text encoding vectors. Then, the vectorized encoded text encoding vector is spliced ​​together with the word encoding vector to form a spliced ​​encoding vector containing global text information and specific word information. It can be understood that the above operation can be performed on multiple user texts to obtain the spliced ​​encoding vectors corresponding to multiple user texts.

[0061] Step S114: Calculate the feature similarity between the concatenated coding vectors, cluster the user texts according to the feature similarity, and obtain multiple clusters.

[0062] The implementation method of the above step S114 is, for example: for the concatenated coding vectors corresponding to every two user texts of the above multiple user texts, the keyword vector Euclidean distance between the two user texts is calculated according to the concatenated coding vectors corresponding to every two user texts to obtain a keyword distance matrix; the average sentence vector Euclidean distance between the two user texts is calculated according to the concatenated coding vectors corresponding to every two user texts to obtain a sentence distance matrix; the text edit distance between the two user texts is calculated according to the concatenated coding vectors corresponding to every two user texts to obtain a text edit distance matrix; according to the keyword distance matrix, the sentence distance matrix and the text edit distance matrix The final distance matrix between the texts is calculated, and the feature similarity between the concatenated coding vector corresponding to the user text and the center vector of the cluster to which it belongs is determined based on the final distance matrix. Then, the similarity matrix is ​​used as the input of the clustering algorithm, and the user texts with similarity greater than the threshold in the similarity matrix are merged into the cluster. The user texts are clustered using clustering algorithms such as the K-means algorithm, the hierarchical clustering algorithm, or the DBSCAN algorithm until the number of the final text set is less than the target cluster, or there is no cluster less than the threshold, so that multiple clusters can be obtained, thereby assigning the user texts to different clusters in the multiple clusters. By calculating the feature similarity between the concatenated coding vectors, the similarity between the texts can be captured more carefully, so that the clustering results are more reasonable, and by calculating the similarity between the concatenated coding vectors, it is easier to identify abnormal texts, which may represent special situations or events.

[0063] Optionally, as another optional implementation of the above step S110, when clustering user texts, clustering can also be performed only based on the text encoding vectors of the user texts. This implementation, for example, includes: vectorizing and encoding the user texts to obtain text encoding vectors, calculating feature similarities between text encoding vectors, clustering the user texts based on feature similarities, and similar user texts are classified into the same cluster, which helps to identify the main topics or emotional tendencies in user feedback.

[0064] Optionally, as another optional implementation of the above step S110, when clustering user texts, clustering can also be performed only based on the word encoding vectors of the text segmentation. For example, the implementation method is as follows: segmenting the user text to obtain multiple words. Vectorizing and encoding multiple words to obtain word encoding vectors. Calculating the feature similarity between the word encoding vectors, clustering the user texts based on the feature similarity so that similar words are classified into the same cluster, which helps to identify the main theme or emotional tendency in user feedback.

[0065] Optionally, after clustering to obtain multiple clusters, for each of the multiple clusters, a preset number of user texts (e.g., the top 5) that best represent the cluster can be screened out from the cluster based on importance metrics (e.g., word frequency, TF-IDF score, or sentiment intensity, etc.), and a pre-trained deep learning model (e.g., BERT or GPT, etc.) can be used to summarize a preset number of user questions from the preset number of user texts. Finally, each user question is compared with the original user text to determine the user text associated with the user question, and the number of user texts corresponding to each user question is counted. If the number of user texts corresponding to a user question is greater than the quantity threshold, an early warning is issued for the user question to identify the most concerned user questions from a large number of user texts. In the above scheme, by counting the number of user texts for each question, an intuitive understanding of the prevalence and urgency of user questions can be obtained, which is of great value for enterprises to improve products, services, or customer experience.

[0066] As another optional implementation of the above step S110, after obtaining multiple words, a word warning may be performed. The implementation of the word warning may include:

[0067] Step S115: for each of the multiple words, obtain the frequency of the word appearing in the user text, and / or count the number of all user texts corresponding to the word.

[0068] For example, the implementation method of the above step S115 is as follows: for each of the multiple words, the frequency of the word appearing in the user text is counted to obtain the frequency of the word appearing in the user text, and the user text in which the word appears can be directly counted to obtain the total number of user texts corresponding to the word. By monitoring the frequency of words, topics or keywords that users are concerned about can be quickly identified. These high-frequency words may be current hot topics or issues of general concern to users. At the same time, counting the number of user texts corresponding to the words can reveal which texts or topics have received more attention, which helps companies quickly understand the focus of user attention. Through the early warning mechanism, companies can quickly notice the problems that frequently appear in user feedback, so as to take timely measures to solve these problems and improve user experience.

[0069] Step S116: If the frequency of the word appearing in the user text is greater than a preset frequency threshold, and / or the number of all user texts corresponding to the word is greater than a preset number threshold, an early warning is issued for the word.

[0070] The implementation method of the above-mentioned step S116 is, for example: the above-mentioned preset frequency threshold can be set according to the specific situation. For example, if a word appears 10 times or more in a user text, it is considered that the word appears more frequently in this text and may be worthy of attention. If a word appears 50 times or more in a user text, it is considered that the word appears very frequently in this text and may need to be focused on. Similarly, the above preset quantity threshold can also be set according to the specific situation, and examples are no longer given here. The above-mentioned warning method for warning words is also called "hot word warning", "word warning" or "soaring word warning", that is, monitoring and alarming for keywords with the same sudden explosion, because this type of warning is based on the warning of real-time data clustering keywords, and there is uncertainty in the problems warned, so it is also called "warning of unknown problems." In the implementation process of the above solution, by using preset category labels, the text can be more accurately classified into the corresponding clusters, which helps to understand the user's needs and problems more deeply. In addition, the use of a large language model for summarization can capture the key information in the text, which helps to understand the core content of user feedback more accurately. By analyzing user feedback from different clusters, companies can discover potential trends and development directions and guide future product and service development.

[0071] As an optional implementation of the above step S120, the above implementation of using the large language model to summarize the questions of all user texts under the cluster may include:

[0072] Step S121: for each of the multiple clusters, label all user texts under the cluster according to a preset category label to obtain multiple text labels.

[0073] As an optional implementation of the above step S121, the above preset category label may include: a product category label and a question category label; the above implementation of labeling all user texts under the cluster according to the preset category label may include: performing structured processing on the user text to obtain structured content. Determine whether the structured content meets the preset conditions, and the preset conditions include: the similarity between the product category in the structured content and the product category label is greater than the first preset similarity threshold, and the similarity between the question category in the structured content and the question category label is greater than the second preset similarity threshold. If the structured content meets the preset conditions, that is, the similarity between the product category in the structured content and the product category label is greater than the first preset similarity threshold, and the similarity between the question category in the structured content and the question category label is greater than the second preset similarity threshold, the question category label can be determined as a text label. Through structured processing and similarity judgment, various aspects of user feedback can be analyzed more carefully, which helps to improve the accuracy and depth of analysis and effectively improves the accuracy of text understanding.

[0074] Step S122: for each text tag among the multiple text tags, use the large language model to perform question summary on the user text corresponding to the text tag to obtain question summary text.

[0075] An implementation of the above-mentioned step S122 is, for example: for each text tag among the multiple text tags, a large language model such as a Transformer-based bidirectional encoder representation (Bidirectional Encoder Representations from Transformers, BERT) model or a bidirectional autoregressive transformer (Bidirectional and Auto-Regressive Transformers, BART) model is used as a text summary model, and a prompt engineering text is generated according to the user text corresponding to the text tag. Then, the prompt engineering text is input into the text summary model, so that the text summary model performs a question summary on the user text corresponding to the text tag according to the prompt engineering text, thereby obtaining a question summary text.

[0076] Before using the text summary model, the prompt engineering text in the training data set can also be used to perform supervised fine-tuning (SFT) training on the text summary model to obtain the above text summary model. The prompt engineering text in the above training data set is as follows:

[0077] Input: You are a language expert and have the following user texts. Please summarize the user's main questions: User text 1; User text 2; User text n.

[0078] Output: Already labeled.

[0079] The training data set may also include user text, a summary question (query) raised in the user text, and a text label corresponding to the user text; wherein the summary question may be a question related to the user text or a question unrelated to the user text.

[0080] Optionally, another implementation of the above step S122 is, for example: using the twin-tower network model trained by the twin-tower network as a large language model, performing a question summary on the user text corresponding to the text label, and obtaining a question summary text. In the training process, the twin-tower network model can input the user text and the candidate summary question, and at the same time input the edit distance, difference text and same text between the user text and the candidate summary question, and then respectively perform attention calculation and training on the user text, the candidate summary question, and the edit distance, difference text and same text between the two, so as to obtain the above-mentioned twin-tower network model. The above-mentioned edit distance (Edit Distance), also known as the Levenshtein distance (Levenshtein Distance), is a measurement method for measuring the degree of difference between two strings. The edit distance defines the minimum number of single-character edit operations required to convert from one string to another. The edit operations here include inserting a character, deleting a character or replacing a character. Specifically, the edit distance can be used to measure the similarity of two strings. If the two strings are exactly the same, their edit distance is 0; if the two strings are completely different, the edit distance is equal to the length of the longest string of the two strings. The above-mentioned difference text refers to the different parts of text between the user text and the candidate summary question, and the above-mentioned same text refers to the same parts of text between the user text and the candidate summary question.

[0081] As an optional implementation of the above step S120, after obtaining multiple text labels, an early warning may be issued for the text labels. The implementation of the early warning for the text labels may include:

[0082] Step S123: for each text tag among the multiple text tags, obtain the total number of user texts corresponding to the text tag.

[0083] It is understandable that the above text labels are determined by the problem category labels. This method of warning text labels is also called label warning, that is, monitoring and warning for abnormal fluctuations in data indicators in the problem category label dimension. Since this type of warning is based on the problem category label dimension, it can also be called "warning for known problems." By obtaining the number of user texts corresponding to each text label, it is possible to quickly identify which topics have received more attention. These topics may be current hot topics or important issues. Through the warning mechanism, companies can quickly notice the problems frequently mentioned in user feedback, so as to take timely measures to solve these problems and improve user experience.

[0084] Step S124: If the number of all user texts corresponding to the text label is greater than a preset number threshold, an early warning is issued for the text label.

[0085] The implementation method of the above steps S123 to S124 is, for example: for each text tag in the multiple text tags, the number of all user texts corresponding to the text tag is counted from the user text database, and if the number of all user texts corresponding to the text tag is greater than a preset number threshold, an early warning is issued for the text tag. The above user text database can be a relational database or a non-relational database, such as a MySQL database or an Oracle database, and a non-relational database such as a Redis database and a MongoDB database.

[0086] Optionally, in addition to monitoring and alerting for abnormal fluctuations in data indicators of the question category label dimension, monitoring and alerting can also be performed for abnormal fluctuations in many indicators, such as: consultation-to-payment order ratio (CPO), negative feedback order volume, negative feedback order rate, etc. Among them, CPO is an important indicator used in the e-commerce industry to measure customer service consultation efficiency and conversion effect. It is calculated by dividing the number of consultations by the total number of payment orders in the same period. If the CPO value of a certain product under each category exceeds 100 and is greater than 110% month-on-month, this is a warning signal, indicating that there are more than 100 consultations behind each payment order on average, and the customer service consultation pressure is much higher than the normal level, which may indicate that customers have encountered more difficulties or doubts in the ordering process. In this case, an abnormal CPO fluctuation warning can be issued for a certain product under this category.

[0087] As an optional implementation of the above step S130, before using the small network model to predict the matching probability between all user texts and question summary texts under the cluster, the small network model may also be trained. The implementation of training the small network model may include:

[0088] Step S131: obtaining a plurality of sample texts, a summary tag and a matching tag of each sample text.

[0089] An example of an implementation of the above step S131 is: the above sample text can be a user text written manually, or a user text automatically generated by a pre-trained large language model. The above summary tag can be a summary tag manually summarizing the user text, or a summary tag automatically summarizing the user text by a text summary model. The above matching tag can be a tag manually annotated after evaluating the user text and the summary tag, or a tag annotated by an executable program through the semantic similarity between the user text and the summary tag, and of course, a tag annotated by a pre-trained large language model through the semantic similarity between the user text and the summary tag.

[0090] Step S132: Use the small network model to predict the matching probability between each sample text and the summary label of the sample text to obtain the predicted probability.

[0091] By training the model to predict the matching probability, the model can learn the deep structure and semantic features of the text and improve its ability to understand the text. This allows the model to learn the semantic consistency between the sample text and the summary label, thereby generating a summary that is more in line with the original meaning.

[0092] Step S133: Calculate the loss value between the predicted probability and the matching label, and train the small network model according to the loss value.

[0093] The implementation method of the above steps S132 to S133 is, for example: constructing a small network model, which can be a deep learning-based architecture, such as a convolutional neural network (CNN) or a long short-term memory network (LSTM) as a small network model, and then using the small network model to predict the matching probability between each sample text and the summary label of the sample text to obtain the predicted probability. The cross entropy loss function is used to calculate the loss value between the predicted probability and the matching label. By calculating the loss value to guide the model training, the model can converge to the optimal state faster and reduce the number of training iterations. In the process of training the small network model according to the loss value, the model parameters of the small network model are updated during the training of the small network model using optimizers such as Adam or SGD for optimization. In the implementation process of the above scheme, by predicting the matching probability, the model can more accurately determine which sample texts match the summary label well, thereby generating a higher quality summary. As the model training proceeds, the model can gradually learn how to generate a summary that is more in line with the content of the sample text, thereby improving the relevance and accuracy of the summary.

[0094] As an optional implementation of the above step S140, the above implementation of determining the final summary result according to the matching probability and the question summary text may include:

[0095] Step S141: Determine whether the matching probability is greater than a preset probability threshold.

[0096] Step S142: If the matching probability is greater than a preset probability threshold, the question summary text is determined as the final summary result.

[0097] In the process from step S141 to step S142, high-quality summaries can be quickly screened out by presetting the probability threshold, avoiding unnecessary re-summarization operations, thereby speeding up the entire summarization process. Only using the large language model for summaries with low matching probabilities can save computing resources, because large language models usually require more computing resources for summarization. Furthermore, a clear probability threshold can help users understand why some summaries are selected and others are reprocessed, increasing users' trust.

[0098] Step S143: If the matching probability is less than or equal to the preset probability threshold, all user texts under the cluster are re-summarized using the large language model to obtain re-summarized text, and the re-summarized text is determined as the final summary result.

[0099] In the process of step S143, for those cases where the quality of the initially generated summaries is poor, re-summarizing using a large language model can effectively correct the errors and improve the robustness of the model. It also allows the model to flexibly adjust its strategy when encountering difficulties, thereby improving the model's ability to cope with summaries of texts of different difficulty levels. With the continuous training and use of the model, the probability threshold can be continuously adjusted to adapt to the new data distribution, reducing the subsequent manual adjustment workload.

[0100] As an optional implementation of the above user text summarization method, after obtaining multiple clusters, a problem summary text warning may also be performed. This implementation may include:

[0101] Step S150: For each cluster among the multiple clusters, if the total number of user texts under the cluster is greater than a preset number threshold, an early warning is issued for the problem summary text of the cluster.

[0102] If a large number of texts are gathered under a certain cluster, it may mean the emergence of a pattern or trend, which helps the system learn and predict future behavior patterns. By setting a quantity threshold to trigger the early warning mechanism, it can promptly detect a sudden increase in attention to a topic or issue, which helps to identify potential important trends or events at an early stage.

[0103] See also Figure 2A schematic diagram of a relationship use case for management using warning rules provided in an embodiment of the present application is shown; optionally, the above-mentioned hot word warnings, tag warnings and problem summary text warnings here can all be managed using warning rules. For example, the warning administrator can add warning rules, modify warning rules, view warning rules (list) and delete warning rules according to the situation, and can also manage warning data, such as viewing warning details information on the warning details page, exporting a table of warning data factors or viewing the original sound (i.e., viewing the original sound data corresponding to the warning user text), etc.

[0104] See also Figure 3 The flowchart of the user text summarization device provided in the embodiment of the present application is shown; the embodiment of the present application provides a user text summarization device 200, including:

[0105] The user text clustering module 210 is used to cluster the user text to be processed to obtain multiple clusters.

[0106] The text question summary module 220 is used to use a large language model to perform question summary on all user texts in each cluster among the multiple clusters to obtain question summary text.

[0107] The matching probability prediction module 230 is used to predict the matching probability between all user texts and question summary texts under the cluster using a small network model, wherein the weight parameter of the small network model is less than the weight parameter of the large language model.

[0108] The summary result determination module 240 is used to determine the final summary result according to the matching probability and the question summary text.

[0109] As an optional implementation of the above device, the user text summarization device further includes:

[0110] The first text extraction submodule is used to extract user text if an order number is identified from return order data, questionnaire survey data, news public opinion data, user evaluation data, online consultation data or customer service work order data.

[0111] The second text extraction submodule is used to extract user text from the user evaluation data, news public opinion data or online consultation data if a product logo is identified from the user evaluation data, news public opinion data or online consultation data.

[0112] As an optional implementation of the above device, the user text clustering module includes:

[0113] The user text segmentation submodule is used to segment the user text to obtain multiple words.

[0114] The word encoding vector acquisition submodule is used to vectorize and encode multiple words to obtain word encoding vectors.

[0115] The word encoding vector concatenation submodule is used to vectorize and encode the user text to obtain a text encoding vector, and concatenate the text encoding vector with the word encoding vector to obtain a concatenated encoding vector.

[0116] The feature similarity clustering submodule is used to calculate the feature similarity between the concatenated encoding vectors and cluster the user texts according to the feature similarity.

[0117] As an optional implementation of the above device, the user text clustering module further includes:

[0118] The word frequency text statistics submodule is used to obtain the word frequency of each word in the user text, and / or to count the number of all user texts corresponding to the word.

[0119] The word warning submodule is used to warn the word if the frequency of the word appearing in the user text is greater than a preset frequency threshold, and / or the number of all user texts corresponding to the word is greater than a preset number threshold.

[0120] As an optional implementation of the above device, the text question summary module includes:

[0121] The text label acquisition submodule is used to label all user texts under each cluster among the multiple clusters according to the preset category label to obtain multiple text labels.

[0122] The label text summary submodule is used to summarize the user text corresponding to each text label in multiple text labels using a large language model.

[0123] As an optional implementation of the above device, the preset category labels include: commodity category labels and question category labels; the text label acquisition submodule includes:

[0124] The structured content obtaining unit is used to perform structured processing on the user text to obtain structured content.

[0125] The structured content judgment unit is used to judge whether the structured content meets preset conditions, wherein the preset conditions include: the similarity between the product category in the structured content and the product category label is greater than a first preset similarity threshold, and the similarity between the question category in the structured content and the question category label is greater than a second preset similarity threshold.

[0126] The text label determination unit is used to determine the question category label as a text label if the structured content meets a preset condition.

[0127] As an optional implementation of the above device, the text question summary module further includes:

[0128] The text quantity obtaining submodule is used to obtain the total user text quantity corresponding to each text tag among the multiple text tags.

[0129] The text tag warning submodule is used to issue a warning for the text tag if the number of all user texts corresponding to the text tag is greater than a preset number threshold.

[0130] As an optional implementation of the above device, the user text summarization device further includes:

[0131] The text label acquisition module is used to acquire multiple sample texts, a summary label and a matching label of each sample text.

[0132] The prediction probability acquisition module is used to use the small network model to predict the matching probability between the summary label of the sample text and each sample text to obtain the prediction probability.

[0133] The loss value calculation module is used to calculate the loss value between the predicted probability and the matching label.

[0134] The network model training module is used to train the small network model according to the loss value.

[0135] As an optional implementation of the above device, the summary result determination module includes:

[0136] The matching probability judgment submodule is used to judge whether the matching probability is greater than a preset probability threshold.

[0137] The summary result determination submodule is used to determine the question summary text as the final summary result if the matching probability is greater than the preset probability threshold; if the matching probability is less than or equal to the preset probability threshold, use the large language model to re-summarize all user texts under the cluster to obtain the re-summary text, and determine the re-summary text as the final summary result.

[0138] As an optional implementation of the above device, the user text summarization device further includes:

[0139] The problem summary warning module is used to issue a warning for the problem summary text of each cluster among the multiple clusters if the total number of user texts under the cluster is greater than a preset number threshold.

[0140] It should be understood that the device corresponds to the above-mentioned user text summarization method embodiment and can execute the various steps involved in the above-mentioned method embodiment. The specific functions of the device can be referred to in the above description, and the detailed description is appropriately omitted here. The device includes at least one software function module that can be stored in a memory in the form of software or firmware or fixed in the operating system (OS) of the device.

[0141] See also Figure 4 The electronic device 300 provided in the embodiment of the present application includes: a processor 310 and a memory 320, wherein the memory 320 stores machine-readable instructions executable by the processor 310, and when the machine-readable instructions are executed by the processor 310, the above method is executed.

[0142] The embodiment of the present application also provides a computer-readable storage medium 330, on which a computer program is stored, and the computer program is executed by the processor 310 to execute the above method. The computer-readable storage medium 330 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, disk or optical disk.

[0143] The embodiment of the present application also provides a computer program product, including: a computer program or a computer instruction, and the computer program or the computer instruction executes the method described above when executed by a processor.

[0144] It should be noted that each embodiment in this specification is described in a progressive manner, and each embodiment focuses on the differences from other embodiments, and the same or similar parts between the embodiments can be referred to each other. For the device embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

[0145] In several embodiments provided by the embodiments of the present application, it should be understood that the disclosed device and method can also be implemented in other ways. The device embodiments described above are only schematic, for example, the flowcharts and block diagrams in the accompanying drawings show the possible implementation architecture, functions and operations of the devices, methods and computer program products according to the multiple embodiments of the embodiments of the present application. In this regard, each box in the flowchart or block diagram can represent a part of a module, a program segment or a code, and a part of a module, a program segment or a code includes one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also be different from the order of occurrence marked in the accompanying drawings. For example, two consecutive boxes can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, which is mainly determined by the functions involved.

[0146] In addition, each functional module of each embodiment in the embodiment of the present application can be integrated together to form an independent part, or each module can exist separately, or two or more modules can be integrated to form an independent part. In addition, in the description of this specification, the description of reference terms "one embodiment", "some embodiments", "example", "specific example", "some examples", etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the embodiment of the present application. In this specification, the schematic representation of the above terms does not necessarily target the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art can combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples without contradiction.

[0147] The above description is only an optional implementation manner of the embodiments of the present application, but the protection scope of the embodiments of the present application is not limited thereto. Any technician familiar with the technical field can easily think of changes or replacements within the technical scope disclosed in the embodiments of the present application, which should be covered within the protection scope of the embodiments of the present application.

Claims

1. A user text summarization method, characterized in that: include: Cluster the user text to be processed to obtain multiple clusters; For each cluster among the multiple clusters, using a large language model to perform question summarization on all user texts under the cluster to obtain a question summary text; Predicting the matching probability between all user texts under the cluster and the question summary text using a small network model, wherein the model parameters of the small network model are less than the model parameters of the large language model; A final summary result is determined according to the matching probability and the question summary text.

2. The method according to claim 1, characterized in that Before clustering the user texts to be processed, the method further includes: If the order number is identified from the return order data, questionnaire survey data, news and public opinion data, user evaluation data, online consultation data or customer service work order data, then the user text is extracted; If the product logo is identified from the user evaluation data, news public opinion data or online consultation data, the user text is extracted from the user evaluation data, news public opinion data or online consultation data.

3. The method according to claim 1, characterized in that: The clustering of the user texts comprises: Performing word segmentation on the user text to obtain multiple words; Performing vectorized encoding on the multiple words to obtain word encoding vectors; Performing vectorization encoding on the user text to obtain a text encoding vector, and concatenating the text encoding vector with the word encoding vector to obtain a concatenated encoding vector; The feature similarity between the concatenated coding vectors is calculated, and the user texts are clustered according to the feature similarity.

4. The method according to claim 3, characterized in that After obtaining the plurality of words, the method further includes: For each of the multiple words, obtain the frequency of occurrence of the word in the user text, and / or count the number of all user texts corresponding to the word; If the frequency of the word appearing in the user text is greater than a preset frequency threshold, and / or the number of all user texts corresponding to the word is greater than a preset number threshold, an early warning is issued for the word.

5. The method according to claim 1, characterized in that The use of the large language model to summarize all user texts under the cluster includes: For each of the multiple clusters, label all user texts under the cluster according to a preset category label to obtain multiple text labels; For each text label among the multiple text labels, a large language model is used to perform question summary on the user text corresponding to the text label.

6. The method according to claim 5, characterized in that The preset category labels include: product category labels and question category labels; and labeling all user texts under the cluster according to the preset category labels includes: Performing structural processing on the user text to obtain structured content; Determining whether the structured content meets a preset condition, wherein the preset condition includes: the similarity between the commodity category in the structured content and the commodity category label is greater than a first preset similarity threshold, and the similarity between the question category in the structured content and the question category label is greater than a second preset similarity threshold; If so, the question category label is determined as the text label.

7. The method according to claim 5, characterized in that After obtaining the plurality of text labels, the method further includes: For each text tag among the multiple text tags, obtaining the total number of user texts corresponding to the text tag; If the total number of user texts corresponding to the text label is greater than a preset number threshold, an early warning is issued for the text label.

8. The method according to claim 1, characterized in that Before using the small network model to predict the matching probability between all user texts under the cluster and the question summary text, the method further includes: Obtaining multiple sample texts, a summary tag and a matching tag of each of the sample texts; Using the small network model, predict the matching probability between each sample text and the summary label of the sample text to obtain a predicted probability; Calculating a loss value between the predicted probability and the matching label; The small network model is trained according to the loss value.

9. The method according to claim 1, characterized in that: The determining of the final summary result according to the matching probability and the question summary text includes: Determine whether the matching probability is greater than a preset probability threshold; If so, the question summary text is determined as the final summary result; otherwise, all user texts under the cluster are re-summarized using the large language model to obtain a re-summary text, and the re-summary text is determined as the final summary result.

10. The method according to any one of claims 1 to 9, characterized in that: After obtaining the plurality of clusters, the method further includes: For each cluster among the multiple clusters, if the total number of user texts under the cluster is greater than a preset number threshold, an early warning is issued for the problem summary text of the cluster.

11. A user text summarization device, characterized in that: include: A user text clustering module is used to cluster the user texts to be processed to obtain multiple clusters; A text question summary module, for summarizing all user texts in each cluster of the multiple clusters using a large language model to obtain a question summary text; A matching probability prediction module, used to predict the matching probability between all user texts under the cluster and the question summary text using a small network model, wherein the weight parameter of the small network model is less than the weight parameter of the large language model; The summary result determination module is used to determine the final summary result according to the matching probability and the question summary text.

12. An electronic device, characterized in that: include: A processor and a memory, wherein the memory stores machine-readable instructions executable by the processor, and the machine-readable instructions are executed by the processor to perform any method according to claims 1 to 10.

13. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method according to any one of claims 1 to 10 is executed.

14. A computer program product, characterized in that include: A computer program or a computer instruction, wherein when the computer program or the computer instruction is executed by a processor, the method according to any one of claims 1 to 10 is executed.

Images