Dialogue abstract generation method and device, computer equipment and storage medium

Abstract

The invention belongs to the field of artificial intelligence and financial science and technology, and relates to a dialogue abstract generation method, which comprises the steps of preprocessing acquired historical dialogue text data, extracting an abstract text, constructing a dialogue sample data set based on the preprocessed dialogue text and the abstract text, and generating a dialogue abstract through the dialogue sample data set. And training a pre-constructed large language model by using LoRa fine tuning to obtain a dialogue abstract generation model, and performing dialogue abstract generation on the to-be-generated dialogue text data by using the dialogue abstract generation model. The invention further provides a dialogue abstract generation device, computer equipment and a storage medium. In addition, the invention also relates to a block chain technology, and the dialogue text data can be stored in a block chain. According to the method, the model training speed and calculation efficiency can be improved, and the abstract generation accuracy and efficiency are improved.

Description

Technical Field

[0001] The present application relates to the fields of artificial intelligence and financial technology, and in particular to a conversation summary generation method, apparatus, computer equipment and storage medium. Background Art

[0002] In today's society, with the improvement of living standards and the diversification of emotional needs, pets have become important members of many families, and people's love and care for pets has reached an unprecedented level. This trend has directly promoted the vigorous development of pet-related service industries. Among them, pet insurance, as an effective means to protect the health of pets and reduce the financial burden of pet owners, has continued to grow in market demand and its market size has expanded day by day.

[0003] However, in the field of pet insurance services, especially in the communication process between customers and insurance companies or service providers, the generation of conversation summaries plays a vital role. It not only helps both parties quickly understand the key points of communication, but also provides important basis for subsequent claims processing, service optimization, etc.

[0004] As the pet insurance market continues to expand, the number of related customer service conversations has increased dramatically. These conversations contain a large amount of key information about pet insurance needs, claims processes, insurance product details, etc. However, traditional conversation summarization methods mainly rely on keyword matching and fixed templates, which show obvious limitations when dealing with complex and changing conversation content. For example, they often fail to effectively identify and extract key information in conversations, and have difficulty adapting to the diversity and dynamic changes of conversation content. In addition, traditional methods are inefficient when processing large amounts of conversation data, and it is difficult to meet the needs of real-time or near real-time summary generation. Summary of the invention

[0005] The purpose of the embodiments of the present application is to propose a conversation summary generation method, apparatus, computer device and storage medium to solve the technical problem that the related technology cannot effectively identify and extract key information in complex and changeable conversations, resulting in low accuracy and efficiency of summary generation.

[0006] In order to solve the above technical problems, the embodiment of the present application provides a method for generating a conversation summary, which adopts the following technical solution:

[0007] Acquire historical conversation text data, preprocess the historical conversation text data to obtain preprocessed conversation text data, and extract summary text corresponding to the preprocessed conversation text data;

[0008] Constructing a conversation sample data set based on the preprocessed conversation text data and the summary text, and randomly dividing the conversation sample data set into a training set and a test set;

[0009] Input the training set into a pre-built large language model and output a predicted summary text;

[0010] Calculate the loss value between the predicted summary text and the annotated summary text based on a preset loss function, and adjust the LoRa parameters in the large language model according to the loss value to obtain a fine-tuning model;

[0011] Evaluate the fine-tuning model using the test set to obtain an evaluation result;

[0012] When the evaluation result meets the expected conditions, outputting the fine-tuned model as a conversation summary generation model;

[0013] The dialogue text data to be generated is obtained, and the dialogue text data to be generated is input into a dialogue summary generation model to generate a target dialogue summary.

[0014] In order to solve the above technical problems, the embodiment of the present application also provides a conversation summary generation device, which adopts the following technical solution:

[0015] An acquisition module, used to acquire historical conversation text data, preprocess the historical conversation text data to obtain preprocessed conversation text data, and extract summary text corresponding to the preprocessed conversation text data;

[0016] A construction module, configured to construct a conversation sample data set based on the preprocessed conversation text data and the summary text, and randomly divide the conversation sample data set into a training set and a test set;

[0017] A prediction module, used for inputting the training set into a pre-built large language model and outputting a predicted summary text;

[0018] An adjustment module, configured to calculate a loss value between the predicted summary text and the annotated summary text based on a preset loss function, and adjust the LoRa parameters in the large language model according to the loss value to obtain a fine-tuning model;

[0019] An evaluation module, used to evaluate the fine-tuning model using the test set to obtain an evaluation result;

[0020] An output module, configured to output the fine-tuned model as a dialogue summary generation model when the evaluation result meets the expected conditions;

[0021] The generation module is used to obtain the dialogue text data to be generated, input the dialogue text data to be generated into the dialogue summary generation model, and generate a target dialogue summary.

[0022] In order to solve the above technical problems, the embodiment of the present application further provides a computer device, which adopts the following technical solution:

[0023] The computer device includes a memory and a processor, wherein the memory stores computer-readable instructions, and the processor implements the steps of the conversation summary generation method as described above when executing the computer-readable instructions.

[0024] In order to solve the above technical problems, the embodiment of the present application further provides a computer-readable storage medium, which adopts the following technical solution:

[0025] The computer-readable storage medium stores computer-readable instructions, and when the computer-readable instructions are executed by the processor, the steps of the conversation summary generation method described above are implemented.

[0026] Compared with the prior art, this application has the following beneficial effects:

[0027] The present application provides a method for generating a conversation summary, which preprocesses the acquired historical conversation text data and extracts the summary text, constructs a conversation sample data set based on the preprocessed conversation text and the summary text, uses the conversation sample data set to fine-tune and train a pre-constructed large language model, obtains a conversation summary generation model, and uses the conversation summary generation model to generate a conversation summary for the conversation text data to be generated; the conversation summary generation model obtained by training in the present application can deeply understand the conversation content, accurately capture key information, and generate a more accurate summary, thereby significantly improving the user experience and service quality, while improving the information processing efficiency; in addition, through the LoRa fine-tuning method, the amount and cost of fine-tuning parameters in the model training process are reduced, the model training speed is improved, and at the same time, a more diverse and comprehensive conversation summary can be generated, thereby improving the accuracy and efficiency of summary generation. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] In order to more clearly illustrate the scheme in the present application, a brief introduction is given below to the drawings required for use in the description of the embodiments of the present application. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0029] Figure 1 is an exemplary system architecture diagram to which the present application may be applied;

[0030] Figure 2 is a flowchart of an embodiment of a method for generating a conversation summary according to the present application;

[0031] Figure 3 yes Figure 2 A flowchart of a specific implementation of step S203;

[0032] Figure 4 is a structural schematic diagram of an embodiment of a conversation summary generating device according to the present application;

[0033] Figure 5 It is a structural diagram of an embodiment of a computer device according to the present application. DETAILED DESCRIPTION

[0034] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by technicians in the technical field of the present application; the terms used in the specification of the application herein are only for the purpose of describing specific embodiments and are not intended to limit the present application; the terms "including" and "having" and any variations thereof in the specification and claims of the present application and the above-mentioned drawings are intended to cover non-exclusive inclusions. The terms "first", "second", etc. in the specification and claims of the present application or the above-mentioned drawings are used to distinguish different objects, not to describe a specific order.

[0035] Reference to "embodiments" herein means that a particular feature, structure, or characteristic described in conjunction with the embodiments may be included in at least one embodiment of the present application. The appearance of the phrase in various locations in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment that is mutually exclusive with other embodiments. It is explicitly and implicitly understood by those skilled in the art that the embodiments described herein may be combined with other embodiments.

[0036] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the accompanying drawings.

[0037] like Figure 1 As shown, the system architecture 100 may include a terminal device 101, a network 102 and a server 103. The terminal device 101 may be a laptop 1011, a tablet computer 1012 or a mobile phone 1013. The network 102 is used to provide a medium for a communication link between the terminal device 101 and the server 103. The network 102 may include various connection types, such as wired, wireless communication links or optical fiber cables.

[0038] The user can use the terminal device 101 to interact with the server 103 through the network 102 to receive or send messages, etc. Various communication client applications can be installed on the terminal device 101, such as web browser applications, shopping applications, search applications, instant messaging tools, email clients, social platform software, etc.

[0039] The terminal device 101 can be any electronic device with a display screen and supporting web browsing. In addition to a laptop computer 1011, a tablet computer 1012 or a mobile phone 1013, the terminal device 101 can also be an e-book reader, an MP3 player (Moving Picture Experts Group Audio Layer III), an MP4 (Moving Picture Experts Group Audio Layer IV), a laptop computer, a desktop computer, etc.

[0040] The server 103 may be a server that provides various services, such as a background server that provides support for a web page displayed on the terminal device 101 .

[0041] It should be noted that the conversation summary generation method provided in the embodiment of the present application is generally executed by a server / terminal device, and accordingly, the conversation summary generation device is generally arranged in the server / terminal device.

[0042] It should be understood that Figure 1 The number of terminal devices, networks and servers in the embodiment is only for illustration. Any number of terminal devices, networks and servers may be provided according to implementation requirements.

[0043] Continue to refer Figure 2 , shows a flowchart of an embodiment of a method for generating a conversation summary according to the present application, comprising the following steps:

[0044] Step S201, obtaining historical conversation text data, preprocessing the historical conversation text data to obtain preprocessed conversation text data, and extracting summary text corresponding to the preprocessed conversation text data.

[0045] In this embodiment, the conversation summary generation method is executed on the electronic device (eg Figure 1 The server / terminal device shown in the figure can obtain the historical conversation text data through a wired connection or a wireless connection. It should be noted that the above wireless connection method may include but is not limited to 3G / 4G / 5G connection, WiFi connection, Bluetooth connection, WiMAX connection, Zigbee connection, UWB (ultra wideband) connection, and other wireless connection methods currently known or to be developed in the future.

[0046] According to the pet insurance business scenario, the historical conversation text data between users and insurance companies is obtained, including conversation records of customer consultation, insurance, claims, etc. The obtained historical conversation text data is preprocessed using natural language processing technology, including word segmentation, stop word removal, special character removal, spelling check and correction, part-of-speech tagging, etc., to obtain standardized text data, namely the preprocessed conversation text data. Through text summarization technology, each preprocessed conversation text is automatically summarized, the key information in the conversation is extracted, and the corresponding summary text is generated.

[0047] For example, first obtain the text of customer conversation records in the past five years from the customer service system of a pet insurance company, totaling 500,000, with an average of at least 20 sentences per conversation. Use word segmentation tools such as jieba, THULAC, FoolNLTK, etc. to segment the text, and use the stop word list to filter out meaningless words, then remove special characters, perform spelling check and correction, and then use the NLTK tool to perform part-of-speech tagging. Finally, use text summarization technology such as the TextRank algorithm and Seq2Seq model to automatically summarize each conversation text, extract a preset number of sentences ranked first as summary text, and calculate the similarity between the summary text and the original text through cosine similarity. If it is lower than the preset threshold, regenerate it, or manually check and correct the generated summary text.

[0048] It should be emphasized that in order to further ensure the privacy and security of the conversation text data, the above conversation text data can also be stored in a blockchain node.

[0049] The blockchain referred to in this application is a new application model of computer technologies such as distributed data storage, peer-to-peer transmission, consensus mechanism, encryption algorithm, etc. Blockchain is essentially a decentralized database, a string of data blocks generated by cryptographic methods. Each data block contains a batch of network transaction information, which is used to verify the validity of its information (anti-counterfeiting) and generate the next block. Blockchain can include the underlying blockchain platform, platform product service layer, and application service layer.

[0050] Step S202: construct a conversation sample data set based on the pre-processed conversation text data and the summary text, and randomly divide the conversation sample data set into a training set and a test set.

[0051] In this embodiment, the preprocessed conversation text data and summary text are converted according to a suitable data structure, for example, stored in the form of conversation-summary pairs, or converted using the JSO structured text format, to facilitate reading and processing of large language models.

[0052] Specifically, a data structure form of a large language model is constructed, and the data structure form includes an instruction part, an input part, and an output part; the summary generation task description is added to the instruction part, the preprocessed dialogue text is added to the input part, and the summary text is added to the output part, so as to form instructions, input text, and output summaries respectively; the instructions, input text, and output summary are concatenated in sequence to form a dialogue sample data set.

[0053] For example, the preprocessed dialogue text and summary text are converted according to the JSON format, which includes the "instruction" field (i.e., the instruction part), the "input" field (input part), and the "output" field (output part). Among them, the "instruction" field is a fixed summary generation task description, such as "generate a summary based on the following dialogue content", the "input" field is the dialogue text, and the "output" field is the corresponding summary. According to the keyword matching rules such as regular expressions, the preprocessed dialogue text and summary text are added to the corresponding fields to form the instructions, input text, and output summary of the corresponding fixed task, and the instructions, input text, and output summary are concatenated into a string in order, separated by special markers, to form a dialogue sample data set that meets the model input requirements. It should be understood that the summary text is equivalent to the real summary label corresponding to the dialogue text in the dialogue sample data set.

[0054] During the conversion process, according to the characteristics and requirements of the summary generation task, the data can be enhanced, such as truncation of dialogue text, length limit of summary, etc., to improve the generalization ability and robustness of the model.

[0055] The converted conversation sample dataset is randomly divided into a training set and a test set in a certain ratio, for example, a ratio of 8:2, for model training, tuning, and evaluation.

[0056] By constructing a conversation sample dataset that meets the input requirements of a large language model, it can be used for subsequent model fine-tuning and optimization to achieve high-quality conversation summary generation, while improving data processing efficiency and model calculation rate.

[0057] Step S203: input the training set into the pre-built large language model and output the predicted summary text.

[0058] In this embodiment, the pre-built large language model is a model that uses the improved ChatGLM2-6b model as the backbone network and is fine-tuned using LoRa.

[0059] In some optional implementations, the pre-built large language model includes an embedding layer, an entity extraction network, a multi-layer encoder introducing LoRa parameters, a multi-layer decoder, and an output layer.

[0060] In this embodiment, see Figure 3 As shown, the step of inputting the training set into the pre-built large language model and outputting the predicted summary text includes:

[0061] Step S301, input the training set into the embedding layer, and convert the dialogue text in the training set into a word vector sequence.

[0062] Among them, the conversation text in the training set is mapped into a low-dimensional, continuous vector space through the embedding layer, which is convenient for the model to process. It can also retain the semantic relationship between the original texts to a certain extent, improve the generalization ability and robustness of the model, and reduce overfitting.

[0063] Step S302, performing entity recognition and semantic relationship extraction on the word vector sequence through an entity extraction network to obtain entity relationship semantic features that integrate context information.

[0064] In this embodiment, the entity extraction network includes a head entity recognition layer, a relationship and tail entity joint recognition layer, and a graph attention layer. The entity extraction network is combined with a pre-built pet insurance domain knowledge graph to perform entity recognition and extraction on the word vector sequence to obtain the key entities in the dialogue text and the semantic relationship between entities.

[0065] Specifically, the word vector sequence is input into the head entity recognition layer for recognition, and all head entities are output; the head entities are associated with each other according to the pre-built knowledge graph to obtain associated entities; the word vector sequence, head entities and associated entities are input into the relation and tail entity joint recognition layer for recognition to obtain all specific relations of each head entity and the corresponding tail entities, and entity relationship triplets are obtained based on the head entity, specific relations and tail entity; the entity relationship triplets are input into the graph attention layer for information transfer and feature fusion to obtain entity relationship semantic features.

[0066] Among them, the word vector sequence h is input through the head entity recognition layer n After identification, the start position (starting position) and end position (ending position) of the entity are obtained to identify all possible head entities. The head entity identification layer uses the following calculation formula:

[0067]

[0068] in, and Respectively represent the word vector sequence h n The probability of the i-th word being recognized as the start position and the end position of the head entity. If the probability exceeds the preset threshold, the tag value of the word is recorded as 1, otherwise it is 0; σ is the sigmoid activation function, ω start With ω endis the trainable weight, b start With b end represents the bias value; and x i =h n [i], represents the word vector sequence h n The i-th marked word in .

[0069] The closest matching principle is: get all and The nearest neighbor The start and The end of 1 is regarded as a complete head entity. This matching method ensures the integrity of the entity and improves the accuracy of head entity recognition.

[0070] Linking the identified head entity with the entity in the pre-built knowledge graph and analyzing the corresponding relationship between the head entity in the knowledge graph can assist in identifying the association between entities in the conversation text, help resolve ambiguity in the text, enable the model to more accurately capture the deep meaning of the text, ensure that the generated summary is consistent with the original text in terms of facts, and reduce the generation of erroneous information.

[0071] After identifying the head entity, the joint identification of the relationship and the tail entity is performed. The structure of each layer of the joint identification layer of the relationship and the tail entity is the same as that of the head entity identification layer. The decoding of the obj entity (tail entity) is similar to the decoding of the sub entity (head entity). The difference between the joint identification layer of the relationship and the tail entity and the head entity identification layer is that: when inputting the word vector sequence, the head entity feature information extracted by the head entity identification layer is considered, and the related entities obtained from the knowledge graph are used as an aid to model the relationship as a function of mapping the head entity to the tail entity.

[0072] In this embodiment, the relationship and tail entity joint identification layer is calculated using the following formula:

[0073]

[0074] in, and Respectively represent the word vector sequence h n The probability of the i-th word being recognized as the start position and the end position of the tail entity. If the probability exceeds the preset threshold, the tag value of the word is recorded as 1, otherwise it is 0; r represents a specific relationship; Represents the average vector of all words between the start and end positions of the k-th head entity.

[0075] It should be noted that when start_o and end_o are both judged as 0 after binary division, it means that there is no corresponding obj entity under this specific relationship, that is, when sub is the head entity, there is no triple of this relationship.

[0076] Next, the attention mechanism of the graph attention layer is used to transfer information and fuse features of the entity-relationship triples, and the entity representation is updated by aggregating the entity's neighbor node information. Each graph attention layer can contain multiple heads, each of which learns a different attention distribution. In the graph attention layer, the features of each node are first mapped to a new feature space through a shared linear transformation (usually a weight matrix), and then a multi-head attention mechanism is used to calculate the weights between the node and its neighbors to obtain entity and relationship representations that fuse contextual information, namely, entity relationship semantic features.

[0077] Through entity extraction, the model can identify key entities in the conversation text and enhance semantic understanding. By capturing the complex relationships between entities through the graph attention layer, it can help the model generate more coherent and consistent summaries and improve the accuracy of the summaries.

[0078] Step S303: Input the entity relationship semantic features into the encoder for encoding, extract the semantic and contextual features of the entity relationship semantic features through the attention mechanism, and obtain the entity relationship feature vector.

[0079] In this embodiment, multiple layers of encoders are stacked, and entity relationship semantic features are encoded in turn through multiple layers of encoders to extract semantic and contextual features of the conversation text. The input of the latter encoder is the output of the previous encoder and the entity relationship semantic features. After being processed by the encoder, a deep semantic feature representation of the conversation text can be obtained, and the output of the last layer of encoders is usually used as the entity relationship feature vector. The encoder introduces LoRa parameters as trainable parameters, and only trains LoRa parameters during the training process.

[0080] Each encoder includes a multi-head self-attention layer, a low-rank adaptation layer, and an encoding feedforward network layer. Specifically, the entity relationship semantic features are input into the multi-head self-attention layer for weight calculation to obtain the entity relationship semantic features; the entity relationship semantic features are input into the low-rank adaptation layer for weight calculation to obtain the entity relationship adaptive features; the entity relationship semantic features and the entity relationship adaptive features are concatenated to obtain the concatenated entity relationship features, and the concatenated entity relationship features are subjected to residual connection and layer normalization processing to obtain the canonical entity relationship features; the canonical entity relationship features are subjected to feature optimization processing through the encoding feedforward network layer to obtain the enhanced entity relationship features; the enhanced entity relationship features are subjected to residual connection and layer normalization processing to obtain the entity relationship feature vector.

[0081] The process of self-attention calculation for each multi-head self-attention layer is as follows:

[0082]

[0083] Among them, Q i =H i-1 W q ; K i =H i-1 W k ; V i =H i-1 W v ;Q i represents the query vector of the i-th head, K i represents the key vector of the i-th head, V i H represents the value vector of the i-th head; i represents the mask-self-attention output of the i-th head; d k Indicates the dimension of the key vector.

[0084] Merging the self-attention, the calculation formula for multi-head attention is as follows:

[0085] MultiHead=Concat(head1,head2,…,head A )W0;

[0086] Among them, Concat represents the matrix concatenation function; W0 represents the parameter matrix when compressing each self-attention.

[0087] In order to better adapt to specific tasks, a low-rank adaptation layer is introduced in the multi-head self-attention layer for fine-tuning. Specifically, the low-rank adaptation layer introduces two low-rank matrices (that is, the rows and columns of the matrix are much less than the rows and columns of the self-attention weight matrix). These two matrices are represented as A and B, respectively, where A is usually the same as the number of rows of the query vector weight matrix, and B is the same as the number of columns of the query vector weight matrix. During inference, the two parts of the results of the multi-head self-attention layer and the low-rank adaptation layer are added and output.

[0088] The encoding feedforward network layer is used to connect and enhance the concatenated entity relationship features obtained by addition to prevent insufficient fitting. Residual connections and layer normalization are performed after each multi-head self-attention layer and encoding feedforward network layer to help gradient flow and reduce degradation problems during training.

[0089] For example, a 6-layer encoder is used. The multi-head self-attention layer of each encoder contains 8 attention heads. The semantic associations between words are extracted from different angles through the multi-head attention mechanism. The LoRa parameters with a rank of 16 are added as trainable parameters in each encoder. The original model parameters are frozen and only the LoRa parameters are trained.

[0090] By introducing LoRa parameters in the encoder, it is possible to capture the key information of the conversation text with fewer parameters, thereby improving the parameter efficiency of the model and reducing the demand for computing resources. It is easier to adapt to summary generation tasks and shows stronger professionalism, thereby improving the accuracy and quality of conversation summaries.

[0091] Step S304: Input the entity relationship feature vector into the decoder, capture the global dependency between the entity relationship feature vectors through the multi-head attention mechanism and the cross attention mechanism, and generate a summary text sequence.

[0092] In this embodiment, multiple layers of decoders are stacked, and entity relationship feature vectors are decoded in sequence by multiple layers of decoders to generate summary text, wherein the decoder generates the summary text by autoregression, and each step predicts the next word based on the previously generated summary text and the output of the encoder. Each layer of encoder includes a position embedding layer, a masked multi-head self-attention layer, an encoder-decoder attention layer, and a decoding feedforward network layer.

[0093] Specifically, the entity relationship feature vector is input into the position embedding layer to obtain the position encoding vector; the position encoding vector and the entity relationship feature vector are input into the masked multi-head self-attention layer for multi-head attention mechanism calculation to obtain the masked attention feature; the masked attention feature is cross-attention mechanism calculated through the encoder-decoder attention layer to obtain the global attention feature; the global attention feature is input into the decoding feedforward network layer for summary generation to obtain the summary text sequence.

[0094] Among them, the position embedding layer is used to provide the position information of words in the entity relationship feature vector, and needs to be combined with the mask (i.e., the masked multi-head attention layer) to ensure that at each step of the generated sequence, the decoder can only see its previous state, but cannot "see the future". The masked multi-head attention layer cannot capture the dependency relationship of words in the following positions because the decoder cannot know future information when decoding, that is, the words in the following positions have not been generated yet. Usually, a Mask matrix is ​​used to mark which positions cannot calculate the attention vector. When calculating the attention vector, the value of the position marked by the Mask matrix will be set to 0; the encoder-decoder attention layer can capture the dependency relationship between any two words and capture contextual information when calculating the attention vector; the decoding feedforward network layer is used to connect and enhance the features output by the encoder-decoder attention layer and generate a summary word vector for the current time step.

[0095] The decoder can process the entire vector in parallel, greatly speeding up summary generation. Through the self-attention mechanism, the decoder can capture long-distance dependencies in the sequence and generate coherent and contextually relevant text. By using mask operations, the decoder cannot see future words when generating each word, which helps avoid error propagation in sequence generation and further improves the accuracy of generated summaries.

[0096] Step S305: perform probability distribution prediction on the summary text sequence through the output layer to obtain a predicted summary text.

[0097] The output layer includes a fully connected layer and a softmax layer. Since the final output of the decoder layer is a summary text sequence, the output summary text sequence is mapped into a longer sequence through the fully connected layer. The text at each position represents the score of the corresponding text. The score is converted into probabilities through the softmax layer, and the preset number of words with the highest probability are output to obtain the probability distribution of each summary text.

[0098] Then, a sequence decoding technology such as Beam Search is used to search and score multiple possible summary texts, and the summary with the highest score is selected as the final output summary text.

[0099] By using a pre-built large language model to predict summary text, it is possible to identify key entities in conversation text, enhance semantic understanding, capture complex relationships between entities, and better extract deep semantic feature representations of conversation text, thereby better capturing long-distance dependencies in sequences, generating coherent and contextually relevant text, and improving the accuracy and efficiency of summary generation.

[0100] Step S204, calculating the loss value between the predicted summary text and the annotated summary text based on a preset loss function, and adjusting the LoRa parameters in the large language model according to the loss value to obtain a fine-tuning model.

[0101] In this embodiment, the preset loss function may adopt a cross entropy loss function, which is used to measure the difference between the probability distribution generated by the model and the probability distribution of the actual next word. The loss value between the predicted summary text and the annotated summary text is calculated by the cross entropy loss function, and the parameters of the low-rank adaptation layer in the encoder are adjusted according to the loss value using the Adam optimizer.

[0102] Step S205, evaluating the fine-tuning model using the test set to obtain an evaluation result.

[0103] Specifically, the test set is input into the fine-tuning model for encoding and decoding to generate a preliminary text summary; the similarity between the preliminary text summary and the corresponding summary text in the test set is calculated to obtain a similarity value, and the similarity value is output as an evaluation result.

[0104] The summary text corresponding to the test set is the real summary. The similarity between the preliminary text summary and the real summary is calculated to evaluate the accuracy and completeness of the summary generation. The similarity can be calculated using cosine similarity, Euclidean distance, Jaccard similarity or edit distance.

[0105] By evaluating the model, we can improve the prediction accuracy and generalization ability of the model, and enhance the stability of the model, thereby improving the quality and efficiency of summary generation.

[0106] Step S206: when the evaluation result meets the expected conditions, the fine-tuned model is output as a dialogue summary generation model.

[0107] In this embodiment, the evaluation result is a similarity value. When the similarity value is greater than or equal to a preset threshold, it means that the model has achieved the expected performance indicators and can generate accurate, complete, high-quality conversation summary text. When the similarity value is less than the preset threshold, the parameters of the fine-tuning model are adjusted, and the feature extraction and summary generation steps are repeated until the generated summary text meets the preset quality requirements.

[0108] Step S207, obtaining the dialogue text data to be generated, inputting the dialogue text data to be generated into the dialogue summary generation model, and generating a target dialogue summary.

[0109] The trained conversation summary generation model is applied to the conversation summary generation task. Specifically, the conversation summary generation model is encapsulated as a Web service using the Flask framework and deployed on a cloud server, realizing the real-time user conversation text summary generation function. After the deployment is completed, the conversation text data to be generated is input into the conversation summary generation model to automatically generate a conversation summary.

[0110] In addition, business-related conversation data can be collected regularly, and the model can be retrained and optimized using incremental learning to adapt to changing business needs, maintaining the high quality and efficiency of conversation summary generation.

[0111] The conversation summary generation model obtained through training in this application can deeply understand the content of the conversation, accurately capture key information, and generate more accurate summaries, thereby significantly improving user experience and service quality, while improving information processing efficiency. In addition, through the LoRa fine-tuning method, the number and cost of fine-tuning parameters in the model training process are reduced, the model training speed is improved, and more diverse and comprehensive conversation summaries can be generated, improving the quality of the summaries. In pet insurance, this can not only enhance the market competitiveness of pet insurance products, but also promote technological progress and innovative development of the entire pet insurance industry.

[0112] The embodiments of the present application can acquire and process relevant data based on artificial intelligence technology. Among them, artificial intelligence (AI) is the theory, method, technology and application system that uses digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results.

[0113] AI basic technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing technology, operation / interaction systems, mechatronics, etc. AI software technologies mainly include computer vision technology, robotics technology, biometrics technology, speech processing technology, natural language processing technology, and machine learning / deep learning.

[0114] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through computer-readable instructions, and the computer-readable instructions can be stored in a computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, the aforementioned storage medium can be a non-volatile storage medium such as a disk, an optical disk, a read-only memory (ROM), or a random access memory (RAM).

[0115] It should be understood that, although the steps in the flowchart of the accompanying drawings are displayed in sequence as indicated by the arrows, these steps are not necessarily executed in sequence in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least a part of the steps in the flowchart of the accompanying drawings may include multiple sub-steps or multiple stages, and these sub-steps or stages are not necessarily executed at the same time, but can be executed at different times, and their execution order is not necessarily sequential, but can be executed in turn or alternately with other steps or at least a part of the sub-steps or stages of other steps.

[0116] Further references Figure 4 , as a response to the above Figure 2 The present application provides an embodiment of a conversation summary generation device, and the device embodiment is similar to Figure 2 Corresponding to the method embodiment shown, the device can be specifically applied to various electronic devices.

[0117] like Figure 4As shown, the conversation summary generation device 400 described in this embodiment includes: an acquisition module 401, a construction module 402, a prediction module 403, an adjustment module 404, an evaluation module 405, an output module 406 and a generation module 407. Among them:

[0118] The acquisition module 401 is used to acquire historical conversation text data, preprocess the historical conversation text data to obtain preprocessed conversation text data, and extract summary text corresponding to the preprocessed conversation text data;

[0119] The construction module 402 is used to construct a conversation sample data set based on the pre-processed conversation text data and the summary text, and randomly divide the conversation sample data set into a training set and a test set;

[0120] The prediction module 403 is used to input the training set into a pre-built large language model and output a predicted summary text;

[0121] The adjustment module 404 is used to calculate the loss value between the predicted summary text and the annotated summary text based on a preset loss function, and adjust the LoRa parameters in the large language model according to the loss value to obtain a fine-tuning model;

[0122] The evaluation module 405 is used to evaluate the fine-tuning model through the test set to obtain an evaluation result;

[0123] The output module 406 is used to output the fine-tuning model as a dialogue summary generation model when the evaluation result meets the expected conditions;

[0124] The generation module 407 is used to obtain the dialogue text data to be generated, input the dialogue text data to be generated into the dialogue summary generation model, and generate the target dialogue summary.

[0125] It should be emphasized that in order to further ensure the privacy and security of the conversation text data, the above conversation text data can also be stored in a blockchain node.

[0126] The conversation summary generation device 400 of the present application, through the conversation summary generation model obtained by training, can deeply understand the conversation content, accurately capture key information, and generate a more accurate summary, thereby significantly improving user experience and service quality, while improving information processing efficiency. In addition, through the LoRa fine-tuning method, the amount and cost of fine-tuning parameters in the model training process are reduced, the model training speed is increased, and at the same time, more diverse and comprehensive conversation summaries can be generated, thereby improving the quality of the summaries.

[0127] In some optional implementations, the building module 402 includes:

[0128] A construction submodule, used to construct a data structure form of the large language model, wherein the data structure form includes an instruction part, an input part and an output part;

[0129] An adding submodule, used for adding a summary generation task description to the instruction part, adding the preprocessed dialogue text to the input part, and adding the summary text to the output part, so as to form instructions, input text, and output summary accordingly;

[0130] The splicing submodule is used to splice the instruction, the input text and the output summary in sequence to form a dialogue sample data set.

[0131] By constructing a conversation sample dataset that meets the input requirements of a large language model, it can be used for subsequent model fine-tuning and optimization to achieve high-quality conversation summary generation, while improving data processing efficiency and model calculation rate.

[0132] In some optional implementations, the large language model includes an embedding layer, an entity extraction network, a multi-layer encoder introducing LoRa parameters, a multi-layer decoder and an output layer, and the prediction module 403 includes:

[0133] An embedding submodule, used for inputting the training set into the embedding layer, and converting the dialogue text in the training set into a word vector sequence;

[0134] An entity extraction submodule, used to perform entity recognition and semantic relationship extraction on the word vector sequence through the entity extraction network to obtain entity relationship semantic features that integrate context information;

[0135] An encoding submodule, used for inputting the entity relationship semantic feature into the encoder for encoding, extracting semantic and contextual features of the entity relationship semantic feature through an attention mechanism, and obtaining an entity relationship feature vector;

[0136] A decoding submodule, used for inputting the entity relationship feature vector into the decoder, capturing the global dependency between the entity relationship feature vectors through a multi-head attention mechanism and a cross attention mechanism, and generating a summary text sequence;

[0137] The prediction submodule is used to perform probability distribution prediction on the summary text sequence through the output layer to obtain a predicted summary text.

[0138] By using a pre-built large language model to predict summary text, it is possible to identify key entities in conversation text, enhance semantic understanding, capture complex relationships between entities, and better extract deep semantic feature representations of conversation text, thereby better capturing long-distance dependencies in sequences, generating coherent and contextually relevant text, and improving the accuracy and efficiency of summary generation.

[0139] In some optional implementations of this embodiment, the entity extraction network includes a head entity recognition layer, a relationship and tail entity joint recognition layer, and a graph attention layer, and the entity extraction submodule is further used to:

[0140] Input the word vector sequence into the head entity recognition layer for recognition, and output all head entities;

[0141] Performing relationship association on the head entity according to a pre-built knowledge graph to obtain an associated entity;

[0142] Input the word vector sequence, the head entity and the associated entity into the relationship and tail entity joint recognition layer for recognition, obtain all specific relationships and corresponding tail entities of each head entity, and obtain entity relationship triples based on the head entity, the specific relationship and the tail entity;

[0143] The entity relationship triplet is input into the graph attention layer for information transfer and feature fusion to obtain the entity relationship semantic feature.

[0144] Through entity extraction, the model can identify key entities in the conversation text and enhance semantic understanding. By capturing the complex relationships between entities through the graph attention layer, it can help the model generate more coherent and consistent summaries and improve the accuracy of the summaries.

[0145] In some optional implementations of this embodiment, the encoder includes a multi-head self-attention layer, a low-rank adaptation layer and an encoding feedforward network layer, and the encoding submodule is further used to:

[0146] Inputting the entity relationship semantic features into the multi-head self-attention layer for weight calculation to obtain entity relationship semantic features;

[0147] Inputting the entity relationship semantic features into the low-rank adaptation layer for weight calculation to obtain entity relationship adaptive features;

[0148] splicing the entity relationship semantic feature and the entity relationship adaptability feature to obtain a spliced ​​entity relationship feature, and performing residual connection and layer normalization processing on the spliced ​​entity relationship feature to obtain a standardized entity relationship feature;

[0149] Performing feature optimization processing on the standard entity relationship features through the encoding feedforward network layer to obtain enhanced entity relationship features;

[0150] The enhanced entity relationship features are subjected to residual connection and layer normalization processing to obtain an entity relationship feature vector.

[0151] By introducing LoRa parameters in the encoder, it is possible to capture the key information of the conversation text with fewer parameters, thereby improving the parameter efficiency of the model and reducing the demand for computing resources. It is easier to adapt to summary generation tasks and shows stronger professionalism, thereby improving the accuracy and quality of conversation summaries.

[0152] In this embodiment, the decoder includes a position embedding layer, a masked multi-head self-attention layer, an encoder-decoder attention layer, and a decoding feedforward network layer, and the decoding submodule is further used to:

[0153] Inputting the entity relationship feature vector into the position embedding layer to obtain a position encoding vector;

[0154] Inputting the position encoding vector and the entity relationship feature vector into the masked multi-head self-attention layer to perform multi-head attention mechanism calculation to obtain a masked attention feature;

[0155] Performing a cross-attention mechanism calculation on the masked attention feature through the encoder-decoder attention layer to obtain a global attention feature;

[0156] The global attention feature is input into the decoding feed-forward network layer for summary generation to obtain a summary text sequence.

[0157] The decoder can process the entire vector in parallel, greatly speeding up summary generation. Through the self-attention mechanism, the decoder can capture long-distance dependencies in the sequence and generate coherent and contextually relevant text. By using mask operations, the decoder cannot see future words when generating each word, which helps avoid error propagation in sequence generation and further improves the accuracy of generated summaries.

[0158] In some optional implementations, the evaluation module 405 includes:

[0159] A summary generation submodule, used for inputting the test set into the fine-tuning model for encoding and decoding to generate a preliminary text summary;

[0160] The calculation submodule is used to calculate the similarity between the preliminary text summary and the corresponding summary text in the test set, obtain a similarity value, and output the similarity value as an evaluation result.

[0161] By evaluating the model, we can improve the prediction accuracy and generalization ability of the model, and enhance the stability of the model, thereby improving the quality and efficiency of summary generation.

[0162] To solve the above technical problems, the present application also provides a computer device. Figure 5 , Figure 5 This is a basic structural block diagram of the computer device in this embodiment.

[0163] The computer device 5 includes a memory 51, a processor 52, and a network interface 53 that are interconnected through a system bus. It should be noted that the figure only shows a computer device 5 with a memory 51, a processor 52, and a network interface 53, but it should be understood that it is not required to implement all the components shown, and more or fewer components can be implemented instead. Among them, those skilled in the art can understand that the computer device here is a device that can automatically perform numerical calculations and / or information processing according to pre-set or stored instructions, and its hardware includes but is not limited to microprocessors, application specific integrated circuits (Application Specific Integrated Circuit, ASIC), programmable gate arrays (Field-Programmable Gate Array, FPGA), digital processors (Digital Signal Processor, DSP), embedded devices, etc.

[0164] The computer device may be a computing device such as a desktop computer, a notebook, a PDA, a cloud server, etc. The computer device may interact with a user through a keyboard, a mouse, a remote controller, a touch pad, or a voice control device.

[0165] The memory 51 includes at least one type of readable storage medium, and the readable storage medium includes flash memory, hard disk, multimedia card, card-type memory (for example, SD or DX memory, etc.), random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, magnetic disk, optical disk, etc. In some embodiments, the memory 51 can be an internal storage unit of the computer device 5, such as a hard disk or memory of the computer device 5. In other embodiments, the memory 51 can also be an external storage device of the computer device 5, such as a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (FlashCard), etc. equipped on the computer device 5. Of course, the memory 51 can also include both the internal storage unit of the computer device 5 and its external storage device. In this embodiment, the memory 51 is generally used to store the operating system and various application software installed on the computer device 5, such as computer-readable instructions of the conversation summary generation method, etc. In addition, the memory 51 can also be used to temporarily store various types of data that have been output or are to be output.

[0166] The processor 52 may be a central processing unit (CPU), a controller, a microcontroller, a microprocessor, or other data processing chips in some embodiments. The processor 52 is generally used to control the overall operation of the computer device 5. In this embodiment, the processor 52 is used to run the computer-readable instructions stored in the memory 51 or process data, such as running the computer-readable instructions of the conversation summary generation method.

[0167] The network interface 53 may include a wireless network interface or a wired network interface. The network interface 53 is generally used to establish a communication connection between the computer device 5 and other electronic devices.

[0168] The conversation summary generation model obtained through training can deeply understand the content of the conversation, accurately capture key information, and generate more accurate summaries, thereby significantly improving user experience and service quality, while improving information processing efficiency. In addition, the LoRa fine-tuning method reduces the amount and cost of fine-tuning parameters during model training, increases the speed of model training, and can generate more diverse and comprehensive conversation summaries, improving the quality of the summaries.

[0169] The present application also provides another implementation, namely, providing a computer-readable storage medium, wherein the computer-readable storage medium stores computer-readable instructions, and the computer-readable instructions can be executed by at least one processor to enable the at least one processor to perform the steps of the conversation summary generation method as described above.

[0170] The conversation summary generation model obtained through training can deeply understand the content of the conversation, accurately capture key information, and generate more accurate summaries, thereby significantly improving user experience and service quality, while improving information processing efficiency. In addition, the LoRa fine-tuning method reduces the amount and cost of fine-tuning parameters during model training, increases the speed of model training, and can generate more diverse and comprehensive conversation summaries, improving the quality of the summaries.

[0171] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present application, or the part that contributes to the prior art, can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk), and includes a number of instructions for a terminal device (which can be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in each embodiment of the present application.

[0172] Obviously, the embodiments described above are only some embodiments of the present application, rather than all embodiments. The preferred embodiments of the present application are given in the accompanying drawings, but they do not limit the patent scope of the present application. The present application can be implemented in many different forms. On the contrary, the purpose of providing these embodiments is to make the understanding of the disclosure of the present application more thorough and comprehensive. Although the present application is described in detail with reference to the aforementioned embodiments, for those skilled in the art, it is still possible to modify the technical solutions recorded in the aforementioned specific implementation methods, or to perform equivalent replacement of some of the technical features therein. Any equivalent structure made using the contents of the specification and drawings of this application, directly or indirectly used in other related technical fields, is similarly within the scope of patent protection of this application.

Claims

1. A method for generating a conversation summary, characterized in that: The steps include: Acquire historical conversation text data, preprocess the historical conversation text data to obtain preprocessed conversation text data, and extract summary text corresponding to the preprocessed conversation text data; Constructing a conversation sample data set based on the preprocessed conversation text data and the summary text, and randomly dividing the conversation sample data set into a training set and a test set; Input the training set into a pre-built large language model and output a predicted summary text; Calculate the loss value between the predicted summary text and the annotated summary text based on a preset loss function, and adjust the LoRa parameters in the large language model according to the loss value to obtain a fine-tuning model; Evaluate the fine-tuning model using the test set to obtain an evaluation result; When the evaluation result meets the expected conditions, outputting the fine-tuned model as a conversation summary generation model; The dialogue text data to be generated is obtained, and the dialogue text data to be generated is input into a dialogue summary generation model to generate a target dialogue summary.

2. The method for generating a conversation summary according to claim 1, characterized in that: The step of constructing a conversation sample data set based on the pre-processed conversation text data and the summary text comprises: Constructing a data structure form of the large language model, wherein the data structure form includes an instruction part, an input part, and an output part; Adding the summary generation task description to the instruction part, the preprocessed dialogue text to the input part, and the summary text to the output part, to form instructions, input text, and output summary respectively; The instructions, the input text, and the output summary are concatenated in sequence to form a dialogue sample data set.

3. The method for generating a conversation summary according to claim 1, characterized in that: The large language model includes an embedding layer, an entity extraction network, a multi-layer encoder introducing LoRa parameters, a multi-layer decoder and an output layer. The step of inputting the training set into the pre-built large language model and outputting the predicted summary text includes: Inputting the training set into the embedding layer, converting the dialogue text in the training set into a word vector sequence; Performing entity recognition and semantic relationship extraction on the word vector sequence through the entity extraction network to obtain entity relationship semantic features that integrate context information; Inputting the entity relationship semantic feature into the encoder for encoding, extracting semantic and contextual features of the entity relationship semantic feature through an attention mechanism, and obtaining an entity relationship feature vector; Inputting the entity relationship feature vector into the decoder, capturing the global dependency between the entity relationship feature vectors through a multi-head attention mechanism and a cross attention mechanism, and generating a summary text sequence; The output layer performs probability distribution prediction on the summary text sequence to obtain a predicted summary text.

4. The method for generating a conversation summary according to claim 3, characterized in that: The entity extraction network includes a head entity recognition layer, a relationship and tail entity joint recognition layer, and a graph attention layer. The step of performing entity recognition and semantic relationship extraction on the word vector sequence through the entity extraction network to obtain entity relationship semantic features that integrate context information includes: Input the word vector sequence into the head entity recognition layer for recognition, and output all head entities; Performing relationship association on the head entity according to a pre-built knowledge graph to obtain an associated entity; Input the word vector sequence, the head entity and the associated entity into the relationship and tail entity joint recognition layer for recognition, obtain all specific relationships and corresponding tail entities of each head entity, and obtain entity relationship triples based on the head entity, the specific relationship and the tail entity; The entity relationship triplet is input into the graph attention layer for information transfer and feature fusion to obtain the entity relationship semantic feature.

5. The method for generating a conversation summary according to claim 3, characterized in that: The encoder includes a multi-head self-attention layer, a low-rank adaptation layer and an encoding feedforward network layer. The step of inputting the entity relationship semantic features into the encoder for encoding, extracting the semantic and contextual features of the entity relationship semantic features through the attention mechanism, and obtaining the entity relationship feature vector includes: Inputting the entity relationship semantic features into the multi-head self-attention layer for weight calculation to obtain entity relationship semantic features; Inputting the entity relationship semantic features into the low-rank adaptation layer for weight calculation to obtain entity relationship adaptive features; splicing the entity relationship semantic feature and the entity relationship adaptability feature to obtain a spliced ​​entity relationship feature, and performing residual connection and layer normalization processing on the spliced ​​entity relationship feature to obtain a standardized entity relationship feature; Performing feature optimization processing on the standard entity relationship features through the encoding feedforward network layer to obtain enhanced entity relationship features; The enhanced entity relationship features are subjected to residual connection and layer normalization processing to obtain an entity relationship feature vector.

6. The method for generating a conversation summary according to claim 3, characterized in that: The decoder includes a position embedding layer, a masked multi-head self-attention layer, an encoder-decoder attention layer and a decoding feedforward network layer. The entity relationship feature vector is input into the decoder, and the global dependency between the entity relationship feature vectors is captured by a multi-head attention mechanism and a cross attention mechanism. The step of generating a summary text sequence includes: Inputting the entity relationship feature vector into the position embedding layer to obtain a position encoding vector; Inputting the position encoding vector and the entity relationship feature vector into the masked multi-head self-attention layer to perform multi-head attention mechanism calculation to obtain a masked attention feature; Performing a cross-attention mechanism calculation on the masked attention feature through the encoder-decoder attention layer to obtain a global attention feature; The global attention feature is input into the decoding feed-forward network layer for summary generation to obtain a summary text sequence.

7. The method for generating a conversation summary according to claim 3, characterized in that: The step of evaluating the fine-tuning model through the test set to obtain an evaluation result comprises: Inputting the test set into the fine-tuning model for encoding and decoding to generate a preliminary text summary; The similarity between the preliminary text summary and the corresponding summary text in the test set is calculated to obtain a similarity value, and the similarity value is output as an evaluation result.

8. A conversation summary generating device, characterized in that: include: An acquisition module, used to acquire historical conversation text data, preprocess the historical conversation text data to obtain preprocessed conversation text data, and extract summary text corresponding to the preprocessed conversation text data; A construction module, configured to construct a conversation sample data set based on the preprocessed conversation text data and the summary text, and randomly divide the conversation sample data set into a training set and a test set; A prediction module, used for inputting the training set into a pre-built large language model and outputting a predicted summary text; An adjustment module, configured to calculate a loss value between the predicted summary text and the annotated summary text based on a preset loss function, and adjust the LoRa parameters in the large language model according to the loss value to obtain a fine-tuning model; An evaluation module, used to evaluate the fine-tuning model using the test set to obtain an evaluation result; An output module, configured to output the fine-tuned model as a dialogue summary generation model when the evaluation result meets the expected conditions; The generation module is used to obtain the dialogue text data to be generated, input the dialogue text data to be generated into the dialogue summary generation model, and generate a target dialogue summary.

9. A computer device, characterized in that: The method comprises a memory and a processor, wherein the memory stores computer-readable instructions, and the processor implements the steps of the conversation summary generation method as described in any one of claims 1 to 7 when executing the computer-readable instructions.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer-readable instructions, and when the computer-readable instructions are executed by the processor, the steps of the conversation summary generation method according to any one of claims 1 to 7 are implemented.

Images