Business file interpretation method and device, electronic equipment and storage medium

Abstract

The embodiment of the invention provides a business file interpretation method and device, electronic equipment and a storage medium, belongs to the technical field of artificial intelligence, and is suitable for the field of financial science and technology. The method comprises the steps of obtaining a target business file and a problem text input by a target object, and selecting a target knowledge base from preset knowledge bases according to the target business file; and then performing question interpretation on the question text according to the target knowledge base, the target business file and a preset initial large language model to obtain a candidate interpretation text, and outputting the candidate interpretation text to the target object. And determining a target weight parameter according to feedback information of the target object on the candidate interpretation text. And performing parameter adjustment on the target weight parameter of the initial large language model to obtain a target large language model. And performing question interpretation on the question text according to the target knowledge base, the target business file and the target large language model to obtain a target interpreted text. According to the embodiment of the invention, the file interpretation accuracy can be improved.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, is applicable to the field of financial technology, and in particular, relates to a method and device for interpreting business documents, an electronic device, and a storage medium. Background Art

[0002] In the financial field, when new businesses need to be developed or new relevant regulations appear in the industry involved in existing businesses, relevant professional documents will be generated, such as professional documents on insurance products. Business personnel of financial institutions need to quickly interpret and learn these relevant professional documents to understand the development direction of the industry and adjust business development strategies.

[0003] In the related technology, the existing initial large language model is usually used to directly interpret the relevant professional documents, but the knowledge involved in the relevant professional documents cannot be effectively recognized by the initial large language model, resulting in inaccurate interpretation. Therefore, how to improve the accuracy of document interpretation has become a technical problem that needs to be solved urgently. Summary of the invention

[0004] The main purpose of the embodiments of the present application is to propose a business document interpretation method and device, electronic device and storage medium, aiming to improve the accuracy of document interpretation.

[0005] To achieve the above-mentioned purpose, a first aspect of an embodiment of the present application proposes a business document interpretation method, the method comprising:

[0006] Obtain the target business file and question text input by the target object;

[0007] Selecting a target knowledge base from a preset knowledge base according to the target business document;

[0008] Interpreting the question text according to the target knowledge base, the target business document and a preset initial large language model to obtain a candidate interpretation text;

[0009] Outputting the candidate interpretation text to the target object;

[0010] Determining a target weight parameter according to feedback information of the target object on the candidate interpretation text;

[0011] Adjusting the parameters of the initial large language model according to the target weight parameter to obtain a target large language model;

[0012] The question text is interpreted according to the target knowledge base, the target business file and the target large language model to obtain a target interpreted text.

[0013] In some embodiments, interpreting the question text according to the target knowledge base, the target business document and a preset initial large language model to obtain a candidate interpretation text includes:

[0014] Performing feature vectorization on the target knowledge base to obtain a knowledge base vector;

[0015] Performing feature vectorization on the target business file to obtain a target file vector;

[0016] Performing attention feature fusion on the knowledge base vector and the target file vector to obtain a reference attention fusion vector;

[0017] Performing feature vectorization on the question text to obtain a question vector;

[0018] The reference attention fusion vector and the question vector are subjected to text generation through the initial large language model to obtain the candidate interpretation text.

[0019] In some embodiments, the step of performing feature vectorization on the target business file to obtain a target file vector includes:

[0020] Extracting the target business file to obtain at least two business sub-files; wherein any two adjacent business sub-files overlap;

[0021] Performing feature vectorization on each of the business sub-files to obtain a local vector;

[0022] Feature fusion is performed on each of the local vectors to obtain the target file vector.

[0023] In some embodiments, the step of performing feature vectorization on each of the business sub-files to obtain a local vector includes:

[0024] Performing text extraction on the business sub-file to obtain a target text;

[0025] Extracting position features of the target text to obtain a text position vector;

[0026] Performing text semantic extraction on the business sub-file to obtain a text semantic vector;

[0027] Perform feature fusion according to the text position vector and the text semantic vector to obtain a text fusion feature vector;

[0028] Extracting images from the business sub-file to obtain a target image;

[0029] Extracting position features of the target image to obtain an image position vector;

[0030] Performing image description processing on the target image to obtain an image meaning vector;

[0031] Perform feature fusion according to the image position vector and the image meaning vector to obtain an image fusion feature vector;

[0032] A self-attention calculation is performed on the text fusion feature vector according to the image fusion feature vector to obtain the local vector.

[0033] In some embodiments, outputting the candidate interpretation text to the target object includes:

[0034] Performing authenticity evaluation on the candidate interpretation text according to the target knowledge base and the target business document to obtain an authenticity score;

[0035] Performing a correlation evaluation on the question text and the candidate interpretation text to obtain a correlation degree score;

[0036] The candidate interpretation text is output to the target object according to the relevance score and the authenticity score.

[0037] In some embodiments, outputting the candidate interpretation text to the target object according to the association degree score and the authenticity score includes:

[0038] Integrate and calculate the relevance score and the authenticity score to obtain a target score;

[0039] If the target score is greater than or equal to a preset benchmark threshold, the candidate interpretation text is output to the target object.

[0040] In some embodiments, before outputting the candidate interpretation text to the target object if the target score is greater than or equal to a preset benchmark threshold, the method further includes:

[0041] Performing compliance assessment on the candidate interpretation text to obtain a compliance score;

[0042] Performing weighted processing on the compliance score and the relevance score to obtain a sub-score;

[0043] The authenticity score and the sub-score are numerically multiplied to obtain a reference score; wherein the authenticity score is 0 or 1;

[0044] The target score is updated according to the reference score.

[0045] To achieve the above-mentioned purpose, a second aspect of an embodiment of the present application provides a business document interpretation device, the device comprising:

[0046] An acquisition module is used to acquire a target business file and a question text input by a target object;

[0047] A knowledge base selection module, used to select a target knowledge base from a preset knowledge base according to the target business document;

[0048] A first interpretation module is used to interpret the question text according to the target knowledge base, the target business document and a preset initial large language model to obtain a candidate interpretation text;

[0049] An output module, used for outputting the candidate interpretation text to the target object;

[0050] A parameter determination module, used to determine a target weight parameter according to feedback information of the target object on the candidate interpretation text;

[0051] A model updating module, used for adjusting the parameters of the initial large language model according to the target weight parameter to obtain a target large language model;

[0052] The second interpretation module is used to interpret the question text according to the target knowledge base, the target business file and the target large language model to obtain a target interpreted text.

[0053] To achieve the above objectives, a third aspect of an embodiment of the present application proposes an electronic device, which includes a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the method described in the first aspect is implemented.

[0054] To achieve the above-mentioned purpose, the fourth aspect of an embodiment of the present application proposes a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method described in the first aspect is implemented.

[0055] The business document interpretation method and device, electronic device and storage medium proposed in the present application first obtain the target business document and the relevant question text. Then determine the target knowledge base according to the target business document, and then use the initial large language model to combine the target knowledge base and the target business document to perform a preliminary interpretation of the question text to generate a candidate interpretation text. Subsequently, the candidate interpretation text is output to the target object, and the feedback information of the target object on the candidate interpretation text is obtained. Then adjust the target weight parameter according to the feedback information, so that the model can be dynamically optimized according to actual needs, thereby improving the relevance and accuracy of the file interpretation. Then optimize the initial large language model to the target large language model through parameter adjustment, ensuring that the model can meet the needs of the target object. Finally, use the optimized target large language model to combine the target knowledge base and the target business document to interpret the question text again, and generate a high-quality target interpretation text. The method of this embodiment optimizes the parameters of the large language model that generates the interpretation text by screening the target knowledge base that matches the target business document and promptly responding to the feedback information of the target object, so as to obtain a more accurate interpretation result, thereby improving the accuracy of the file interpretation, and finally providing a more efficient, accurate and reliable document interpretation method for the business personnel of the financial institution, helping them to quickly grasp the changes in the industry and optimize the business development strategy. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 is a flow chart of a business document interpretation method provided by an embodiment of the present application;

[0057] Figure 2 yes Figure 1 Flow chart of step S103 in FIG.

[0058] Figure 3 yes Figure 2 Flow chart of step S202 in FIG.

[0059] Figure 4 yes Figure 1 Flow chart of step S104 in FIG.

[0060] Figure 5 yes Figure 4 Flow chart of step S403 in FIG.

[0061] Figure 6 yes Figure 4 Another optional flowchart of step S403 in;

[0062] Figure 7 It is a structural diagram of a business document interpretation device provided in an embodiment of the present application;

[0063] Figure 8 It is a schematic diagram of the hardware structure of the electronic device provided in the embodiment of the present application. DETAILED DESCRIPTION

[0064] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0065] It should be noted that, although the functional modules are divided in the device schematic diagram and the logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than the module division in the device or the order in the flowchart. The terms "first", "second", etc. in the specification, claims and the above drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence.

[0066] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art to which this application belongs. The terms used herein are only for the purpose of describing the embodiments of this application and are not intended to limit this application.

[0067] First, some nouns involved in this application are analyzed:

[0068] Artificial intelligence (AI) is a new technical science that studies and develops theories, methods, technologies and application systems for simulating, extending and expanding human intelligence. AI is a branch of computer science that attempts to understand the essence of intelligence and produce a new type of intelligent machine that can respond in a similar way to human intelligence. Research in this field includes robotics, speech recognition, image recognition, natural language processing and expert systems. AI can simulate the information process of human consciousness and thinking. AI is also a 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.

[0069] Natural language processing (NLP): NLP uses computers to process, understand and apply human languages ​​(such as Chinese, English, etc.). NLP is a branch of artificial intelligence and an interdisciplinary subject between computer science and linguistics. It is often referred to as computational linguistics. Natural language processing includes grammatical analysis, semantic analysis, and text understanding. Natural language processing is often used in technical fields such as machine translation, handwritten and printed character recognition, speech recognition and text-to-speech conversion, information intent recognition, information extraction and filtering, text classification and clustering, public opinion analysis and opinion mining. It involves data mining related to language processing, machine learning, knowledge acquisition, knowledge engineering, artificial intelligence research, and linguistic research related to language computing.

[0070] Information Extraction: A text processing technology that extracts specified types of entity, relationship, event and other factual information from natural language text and forms structured data output. Information extraction is a technology that extracts specific information from text data. Text data is composed of some specific units, such as sentences, paragraphs, and chapters. Text information is composed of some small specific units, such as characters, words, phrases, sentences, paragraphs, or a combination of these specific units. Extracting noun phrases, names, place names, etc. from text data are all text information extraction. Of course, the information extracted by text information extraction technology can be various types of information.

[0071] Image description (ImageCaption) generates a natural language description for an image and uses the generated description to help applications understand the semantics expressed in the image visual scene. For example, image description can convert image retrieval into text retrieval, which is used to classify images and improve image retrieval results. People usually only need to take a quick glance to describe the details of the image visual scene, while automatically adding descriptions to images is a comprehensive and arduous computer vision task that requires converting the complex information contained in the image into a natural language description. Compared with ordinary computer vision tasks, image captioning requires not only identifying objects from images, but also associating the identified objects with natural semantics and describing them in natural language. Therefore, image description requires people to extract deep features of images, associate with semantic features and convert them for generating descriptions.

[0072] In the financial field, when there is a need to develop new business or new relevant regulations appear in the industry involved in existing business, relevant professional documents will be generated. The business personnel of financial institutions need to quickly interpret and study these relevant professional documents to understand the development direction of the industry and adjust their business development strategies.

[0073] In the related art, the existing initial large language model is usually used to directly interpret the relevant professional documents, but the knowledge involved in the relevant professional documents cannot be effectively recognized by the initial large language model, resulting in inaccurate interpretation. When the file to be processed is a large file, the performance requirements for interpretation are high, which may cause the model interpretation to take a long time or even fail. Therefore, how to improve the accuracy of file interpretation has become a technical problem that needs to be solved urgently.

[0074] Based on this, the embodiments of the present application provide a business document interpretation method and device, an electronic device and a storage medium, aiming to improve the accuracy of document interpretation.

[0075] The business document interpretation method and device, electronic device and storage medium provided in the embodiments of the present application are specifically described through the following embodiments. First, the business document interpretation method in the embodiments of the present application is described.

[0076] 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.

[0077] 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.

[0078] The business file interpretation method provided in the embodiment of the present application relates to the field of artificial intelligence technology. The business file interpretation method provided in the embodiment of the present application can be applied to the terminal, can also be applied to the server side, and can also be software running in the terminal or the server side. In some embodiments, the terminal can be a smart phone, a tablet computer, a laptop computer, a desktop computer, etc.; the server side can be configured as an independent physical server, or it can be configured as a server cluster or distributed system composed of multiple physical servers, and can also be configured as a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms; the software can be an application that implements the business file interpretation method, etc., but is not limited to the above forms.

[0079] The present application can be used in many general or special computer system environments or configurations. For example: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, distributed computing environments including any of the above systems or devices, etc. The present application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc. that perform specific tasks or implement specific abstract data types. The present application can also be practiced in distributed computing environments, in which tasks are performed by remote processing devices connected through a communication network. In a distributed computing environment, program modules can be located in local and remote computer storage media including storage devices.

[0080] Figure 1 is an optional flowchart of the business document interpretation method provided in the embodiment of the present application. Figure 1 The method may include but is not limited to steps S101 to S107.

[0081] Step S101, obtaining a target business file and a question text input by a target object.

[0082] Step S102: selecting a target knowledge base from preset knowledge bases according to the target business document.

[0083] Step S103, interpreting the question text according to the target knowledge base, the target business document and the preset initial large language model to obtain candidate interpretation texts.

[0084] Step S104: output the candidate interpretation text to the target object.

[0085] Step S105, determining a target weight parameter according to the feedback information of the target object on the candidate interpretation text.

[0086] Step S106, adjusting the parameters of the initial large language model according to the target weight parameter to obtain a target large language model.

[0087] Step S107, interpreting the question text according to the target knowledge base, the target business file and the target large language model to obtain a target interpreted text.

[0088] Steps S101 to S107 shown in the embodiment of the present application are first performed by obtaining target business documents and relevant question texts. Then the target knowledge base is determined according to the target business documents, and then the initial large language model is used to combine the target knowledge base and the target business documents to perform a preliminary interpretation of the question text to generate a candidate interpretation text. Subsequently, the candidate interpretation text is output to the target object, and the feedback information of the target object on the candidate interpretation text is obtained. Then the target weight parameter is adjusted according to the feedback information so that the model can be dynamically optimized according to actual needs, thereby improving the relevance and accuracy of file interpretation. Then the initial large language model is optimized to the target large language model through parameter adjustment, ensuring that the model can meet the needs of the target object. Finally, the optimized target large language model is used to combine the target knowledge base and the target business documents to interpret the question text again to generate a high-quality target interpretation text. The method of this embodiment optimizes the parameters of the large language model that generates the interpretation text by screening the target knowledge bureau that matches the target business document and promptly responding to the feedback information of the target object, so as to obtain more accurate interpretation results, thereby improving the accuracy of document interpretation, and ultimately providing the business personnel of the financial institution with a more efficient, accurate and reliable document interpretation method, helping them to quickly grasp industry changes and optimize business development strategies.

[0089] In step S101 of some embodiments, the target object is a user group that needs to interpret and learn the file. The target business file is a business-related file that the target object currently needs to learn and understand, and its form can be text, pictures, videos, and audio, etc., but is not limited to this. The question text is the question that the target object has about the file. In the application scenario of the present embodiment where a new business is launched or a new specification appears in the current business, the target object can be a business processing personnel of a financial institution, and the target business file can be relevant legal and regulatory documents, industry standards, business process documents, operation manuals, risk assessment reports, etc.

[0090] In some embodiments, the target object can input the target business file through a mobile application, a software development kit (SDK), a web page, etc.

[0091] In some embodiments, the target business file needs to be preprocessed after it is input, including format conversion and text cleaning, to ensure the consistency of the input data and reduce interference information, laying the foundation for the model to efficiently interpret the file.

[0092] In step S102 of some embodiments, the professional field of the target business document can be directly identified by a pre-trained large language model, or the target object can directly input the professional field of the target business document. Then, a knowledge base matching the professional field is selected from the preset knowledge base as the target knowledge base.

[0093] In some embodiments, the preset knowledge base includes not only documents publicly available on the Internet, but also internal documents of the financial institution where the target object is located, such as historical transaction contracts, business processes, compliance requirements and risk management strategies, etc. These knowledge bases can provide normative guidance for document interpretation.

[0094] In step S103 of some embodiments, a large language model technology may be used to interpret the question. For details, see Figure 2 In some embodiments, step S103 may include but is not limited to steps S201 to S205:

[0095] Step S201, feature vectorization is performed on the target knowledge base to obtain a knowledge base vector.

[0096] Step S202, feature vectorization is performed on the target business file to obtain a target file vector.

[0097] Step S203, performing attention feature fusion on the knowledge base vector and the target file vector to obtain a reference attention fusion vector.

[0098] Step S204, feature vectorization is performed on the question text to obtain a question vector.

[0099] Step S205, generating text from the reference attention fusion vector and the question vector through the initial large language model to obtain candidate interpretation text.

[0100] In step S201 of some embodiments, key text information needs to be extracted from the target knowledge base, which may be the text content of each knowledge item in the knowledge base, the title of a knowledge point, important concepts, etc. Then, using the word embedding method in natural language processing technology, such as the Word2Vec algorithm and the GloVe algorithm, each word in the extracted key text information is mapped to a vector representation of a fixed dimension. Then, the vectors of each word can be integrated by methods such as average pooling and maximum pooling, and finally a vector that can represent the characteristics of the entire target knowledge base, that is, the knowledge base vector, is obtained.

[0101] In step S202 of some embodiments, the target file vector is a vector representation of the target business file. In some embodiments, the feature vector of the target business file can be obtained by the same method as the embodiment of step S201. In other embodiments, the target business file is long and requires more content to be interpreted, so the understanding of the context is high, resulting in high difficulty in interpreting such files. For this, please refer to Figure 3Step S202 may include but is not limited to steps S301 to S303:

[0102] Step S301: extract the target business file to obtain at least two business sub-files.

[0103] Step S302, perform feature vectorization on each business sub-file to obtain a local vector.

[0104] Step S303, performing feature fusion on each local vector to obtain a target file vector.

[0105] In step S301 of some embodiments, the business subfile refers to the independent local content of the target business file, and there is overlap between two adjacent business subfiles at any position in order to preserve the semantic coherence of the previous and next subfiles.

[0106] In step S302 of some embodiments, there are text and images in the target business file at the same time, and step S302 includes feature processing of text data and image data. Specifically, firstly, text extraction is performed on the business sub-file to obtain the target text. Then, position feature extraction is performed on the target text to obtain a text position vector. This step extracts features based on the sequential position information of the target text in the entire target business file. For example, a position index can be assigned to each word or character in the target text, and these position indexes are converted into vector form through a suitable encoding method, such as One-Hot Encoding, or the embedding layer in the deep learning network framework can be used to map the position index to a vector. After the position indexes of all words are mapped, text position vectors are obtained, and these vectors can reflect the position features of each element in the text in sequence. Subsequently, text semantics extraction is performed on the business sub-file to obtain a text semantic vector. This step can use the BERT (Bidirectional Encoder Representations from Transformers) model. First, preprocess the target text according to the format required by the model (such as word segmentation, adding specific tags, etc.), and then input the processed text into the BERT model. The BERT model will generate a corresponding semantic representation vector for each word in the text based on its semantic understanding ability obtained through pre-training on large-scale corpora, and finally integrate all vectors to obtain the text semantic vector of the business sub-file. Then, feature fusion is performed based on the text position vector and the text semantic vector to obtain a text fusion feature vector. Feature fusion can be performed by directly concatenating two vectors or by adding the two vectors element by element.

[0107] For the image data of the business sub-file, firstly, the business sub-file is subjected to image extraction to obtain the target image. Then, the position feature extraction is performed on the target image to obtain the image position vector. The image position vector characterizes the position feature of the target image in the business sub-file. Subsequently, the target image is subjected to image description processing to obtain the image meaning vector. Exemplarily, a pre-trained image description model is used to extract the features of the target image, and then a natural language text describing the image content is generated by a generative language model based on the extracted features, and then this text is converted into a vector form by the aforementioned text semantic extraction method, and this vector can be used as the image meaning vector. In addition, a method combining image classification and semantic embedding can also be adopted, firstly, the pre-trained image classification model is used to classify the target image to obtain the category label (for example, the category of landscape image, figure image, etc.) to which the image belongs, and then the pre-trained word embedding model is used to convert the words corresponding to the category label into vectors, thereby obtaining the image meaning vector, so as to mine the general meaning information of the image from the perspective of image classification and embody it in the form of a vector. Finally, feature fusion is performed according to the image position vector and the image meaning vector to obtain the image fusion feature vector. The method of realizing feature fusion is the same as that in the above embodiment, but it is only used as an example and does not represent a strict limitation to this embodiment.

[0108] Finally, the text fusion feature vector is self-attentioned based on the image fusion feature vector to obtain the local vector. The self-attention mechanism can automatically capture the correlation and importance between different elements in the text fusion feature vector and the image fusion feature vector. Whether it is text or image data, there are often complex semantic and structural relationships within and between them. The self-attention mechanism can flexibly learn and adjust the focus according to the specific content of the data and task requirements. For example, in some scenarios, a specific area in the image is closely related to several keywords in the text. The self-attention mechanism can capture this relationship and give it appropriate reflection in the calculation, which is finally reflected more accurately in the local vector.

[0109] Exemplarily, three linear transformation layers can be constructed in a deep learning framework, and the text fusion feature vector and the image fusion feature vector are respectively passed through three linear transformation layers to obtain the corresponding query vector (Query Vector), key vector (Key Vector) and value vector (Value Vector), and the query vector corresponding to the image fusion feature vector and the key vector of the text fusion feature vector are similarly calculated and normalized to obtain the attention weight representing the correlation between the two, and the value vector corresponding to the text fusion feature vector is weightedly summed according to the weight to obtain a local vector.

[0110] In step S303 of some embodiments, fusion can be achieved through feature concatenation, weighted fusion according to preset weight coefficients, and self-attention calculation, and the obtained target file vector is a feature vector representation of the target business file.

[0111] Steps S301 to S303 shown in the embodiment of the present application can extract at least two business sub-files from the target business file, which can break down the originally complex, large and possibly complicated business file into different sub-files, and the presence of duplicate content between every two adjacent sub-files can avoid missing key content when processing large-scale complex files, and ensure the coherence of the final contextual understanding. Then, feature vectorization is performed on each business sub-file to obtain a local vector, and finally, feature fusion is performed on each local vector to obtain a target file vector, forming a vector that can fully reflect the overall characteristics of the original target business file, which helps the large language model question-answering system to accurately interpret question texts in combination with business files, generate high-quality candidate interpretation texts, and other aspects, thereby improving the accuracy of the output results.

[0112] In step S203 of some embodiments, an attention mechanism can be used to achieve fusion. The attention mechanism enables the model to automatically focus on the key parts of the input information. In this scenario, the similarity between the knowledge base vector and the target file vector is calculated. Then, different weights are assigned to different elements in the two vectors according to the similarity, and higher weights are given to elements with strong correlation, and lower weights are given to elements with strong correlation. Finally, the elements of the two vectors are weighted and summed according to these weights to obtain a new vector, which is the reference attention fusion vector. This vector combines the feature information in the knowledge base and the target business file that is considered to be critical and interrelated under the attention mechanism. For example, if a dimension in the knowledge base vector has a high correlation with the corresponding dimension in the target file vector, the dimension will have a larger weight to participate in the calculation during fusion, so that the fused vector will focus more on reflecting this highly correlated feature.

[0113] In step S204 of some embodiments, the question vector is a vector obtained by processing the question text in natural language form through feature vectorization, which embodies the key features of the question text such as semantics and grammar in a numerical form. Exemplarily, a word embedding algorithm can be used to convert the words in the question text into vectors. For the entire question sentence, the word vector can be further encoded by using a network structure such as a recursive neural network and a long short-term memory network (LSTM) that can process sequence information, and finally a vector that can represent the overall semantic features of the question text, i.e., the question vector, is generated.

[0114] In step S205 of some embodiments, the initial large language model is a pre-trained language model. The candidate interpretation text needs to provide the reference attention fusion vector and the question vector as input to the initial large language model. The initial large language model uses the language knowledge learned during pre-training and the word combination, semantic coherence and other rules learned from a large amount of text data based on the input vector to analyze the semantics of the question, the associated knowledge base and business file information from the given vector information, and then gradually generates the corresponding text content, and finally forms multiple possible candidate interpretation texts.

[0115] Steps S201 to S205 shown in the embodiment of the present application firstly feature vectorize the target knowledge base and the target business file respectively, so as to convert the knowledge and business-related content originally in text form into a computable and easy-to-analyze vector form. Then, the knowledge base vector and the target file vector are fused through attention features, so that the key information of the two can be dynamically weighted and integrated according to their relevance, so that the reference attention fusion vector covers both the professional knowledge of the knowledge base and the actual situation of the business file, providing a more targeted and business-oriented information basis for subsequent answering questions. Then, the question text is also feature vectorized, and finally, the text is generated based on the fusion vector and the question vector with the help of the initial large language model, which can make full use of the fusion features integrated in the previous steps, so that the candidate interpretation text generated by the large language model is more in line with the question intent and more in line with the business background knowledge.

[0116] In step S104 of some embodiments, it is necessary to perform multi-dimensional scoring on the candidate interpretation texts output by the initial large language model, evaluate the reliability of the current candidate interpretation texts according to the scores of different dimensions, and push the candidate interpretation texts that are reliable to the target object according to the scores. The evaluation dimensions for the candidate interpretation texts include but are not limited to authenticity and relevance. For details, please refer to Figure 4 In some embodiments, step S104 may include but is not limited to steps S401 to S403:

[0117] Step S401 , performing authenticity evaluation on the candidate interpretation text according to the target knowledge base and the target business document to obtain an authenticity score.

[0118] Step S402, evaluating the relevance of the question text and the candidate interpretation texts to obtain a relevance score.

[0119] Step S403: output the candidate interpretation text to the target object according to the relevance score and the authenticity score.

[0120] In step S401 of some embodiments, authenticity refers to whether the content represented by the candidate interpretation text is authentic and reliable, which is specifically manifested in whether the logic before and after the text content is contradictory, and whether there is a statement in the text that violates the target knowledge base. For the implementation of authenticity evaluation, the logic of the candidate interpretation text can be verified by a pre-trained deep learning network. The final authenticity score can be in the range of [0,1].

[0121] In step S402 of some embodiments, the degree of association refers to the correlation between the candidate interpretation text and the question text, which can be used to determine whether the candidate interpretation text is closely centered around the question and whether it can effectively respond to the target object's questions, thereby screening out interpretation texts that are closely related to the question. For the implementation of the degree of association evaluation, the words in the question text and the candidate interpretation text can be converted into word vector forms using word embedding technology. Then, the similarity between the word vectors of the two is calculated, and the similarity can be measured by calculating indicators such as cosine similarity or Euclidean distance. The similarity data is used as a degree of association score. It should be noted that the value range of the degree of association score needs to be consistent with the value range of the authenticity score, which can be [0,1].

[0122] In step S403 of some embodiments, please refer to Figure 5 Step S403 may also include but is not limited to steps S501 to S502:

[0123] Step S501, integrating and calculating the relevance score and the authenticity score to obtain a target score.

[0124] Step S502: if the target score is greater than or equal to a preset reference threshold, the candidate interpretation text is output to the target object.

[0125] In step S501 of some embodiments, the target score is a comprehensive quantitative indicator of the performance of the candidate interpretation text in different evaluation dimensions. The target score can be calculated by pre-assigning corresponding weight data to the relevance score and the authenticity score, and performing weighted summation of the relevance score and the authenticity score according to the weight data, but is not limited thereto.

[0126] In step S502 of some embodiments, the reference threshold may be 0.8, but is not limited thereto. If the target score is less than the reference threshold, the candidate interpretation text corresponding to the target score will be eliminated and will not be pushed to the target object.

[0127] Steps S501 to S502 shown in the embodiment of the present application integrate the relevance score and the authenticity score to calculate the target score, thereby comprehensively considering the pros and cons of the candidate interpretation text from multiple key evaluation dimensions and avoiding the one-sidedness of single-dimensional evaluation. At the same time, the decision on whether to output the candidate interpretation text is based on the preset benchmark threshold, which further ensures that the output content meets certain standards in terms of relevance and authenticity, and helps to improve the overall quality of the interpretation text finally presented to the target object, making it more in line with the problem requirements and more in line with the actual situation, thereby increasing the probability of the target object obtaining accurate and useful information, thereby improving user satisfaction.

[0128] In some other embodiments, the compliance evaluation dimension of the candidate interpretation document may be increased. In the multi-dimensional comprehensive evaluation scheme, the authenticity score may be set as a decisive influencing factor. Figure 6 Before step S502, the business file interpretation method provided in the embodiment of the present application further includes but is not limited to steps S601 to S604:

[0129] Step S601, conduct compliance assessment on the candidate interpretation text to obtain a compliance score.

[0130] Step S602: weighting the compliance score and the relevance score to obtain a sub-score.

[0131] Step S603, performing numerical multiplication calculation on the authenticity score and the sub-score to obtain a reference score.

[0132] Step S604: update the target score according to the reference score.

[0133] In step S601 of some embodiments, compliance means that the content of the candidate interpretation text should not involve violence, illegal or irregular content, insulting words, negative values, etc.

[0134] Specifically, a list of blocked words can be constructed, the candidate interpretation texts can be segmented, each word can be compared in the blocked word list, the number of hit keywords can be counted, and the number of hit keywords can be compared with the total number of words in the candidate interpretation text. The compliance score can be obtained according to the ratio. Exemplarily, the candidate interpretation text is decomposed into 10 words, of which there is 1 hit keyword, then the compliance score can be 1-(1 / 10)=0.9. The value range of the compliance score is the same as the value range of the scores of other dimensions, which can be [0,1]. In other embodiments, a language model that has been trained with a large number of compliant and non-compliant text pairs can be selected, and the candidate interpretation text can be input into the model. The model predicts the compliance probability based on the learned semantic and contextual features, and maps the probability value to a compliance score, that is, when the probability is 0.8, the corresponding compliance score is also 0.8.

[0135] In step S602 of some embodiments, the sub-score combines the compliance score and the relevance score of the candidate interpretation text, and obtains a comprehensive quantitative index after weighted processing. Specifically, corresponding weight data is assigned to the compliance score and the relevance score in advance, and then the relevance score and the authenticity score are weighted and summed according to the weight data to obtain the sub-score.

[0136] In step S603 of some embodiments, the authenticity score is set to 0 or 1. Once there is a logical inconsistency in the candidate interpretation text, or any statement is contrary to the target business document, the target knowledge base and common knowledge, the authenticity score is directly determined to be 0, and then the target score of the corresponding candidate interpretation text is directly set to zero. If the candidate interpretation text does not have the above situation, the authenticity score is directly determined to be 1. It can be understood that the reference score can satisfy the following analytical formula:

[0137] M0=M t ·(w1·M r +w2·M c ) (1),

[0138] Among them, M0 represents the reference score. t Represents the authenticity score, which can be 0 or 1. r Indicates the correlation score. c Indicates the compliance score. w1 indicates the weight data corresponding to the association degree score, and w2 indicates the weight data corresponding to the compliance score.

[0139] In step S604 of some embodiments, the reference score is used as the target score to participate in the judgment of the subsequent step S502.

[0140] Steps S601 to S604 shown in the embodiment of the present application add a compliance evaluation dimension to the original evaluation system, binary the authenticity score, ensure the reliability of the candidate interpretation text, and then calculate the score data of the three dimensions to obtain a reference score. Finally, the target score is updated according to the reference score, which helps to push more accurate text content to the target object.

[0141] In steps S401 to S403 shown in the embodiment of the present application, the authenticity of the candidate interpretation text is first evaluated based on the target knowledge base and the target business file to ensure the reliability of the output content in reflecting the actual business situation, and then the degree of relevance to the question text is measured through relevance evaluation to avoid irrelevant answers. Finally, the candidate interpretation text is output based on a comprehensive multi-dimensional score, so that the presented content is both in line with the core of the question and has factual basis, which improves the quality of information provided to the target object in the question-answering interaction process as a whole, helps the target object to obtain useful interpretation content more efficiently and accurately, and improves the practicality and user experience of the entire question-answering system.

[0142] In step S105 of some embodiments, feedback information refers to the evaluation, opinions, supplementary explanations, etc. of the candidate interpretation text provided by the target object through the interactive interface after viewing the candidate interpretation text. Specifically, the embodiment of the present application provides a corresponding feedback collection mechanism to obtain the feedback information of the target object for the candidate interpretation text. For example, an evaluation button is set on the interactive interface, such as "satisfied", "unsatisfied", and "partially correct". Or the target object enters specific modification opinions, supplementary explanations, etc. of the currently pushed interpretation text through the text input box. Then, the dimension that the target object is dissatisfied with is determined according to the feedback information, so as to determine the weight parameter that the initial large language model needs to adjust, that is, the target weight parameter. Exemplarily, when the target user feedbacks that the answer content is too divergent, the target weight parameter can be a temperature parameter. When the feedback information indicates that the answer content is not related to the question, the target weight parameter can be a related sampling parameter, such as the parameter top_k and the parameter top_p. When the feedback information indicates that the answer content is incoherent, the target weight parameter can be a maximum sampling length parameter.

[0143] It is understandable that when the target object is satisfied with the candidate interpretation text, he can give feedback on the corresponding evaluation, or directly enter the next question text without feedback operation. At this time, no parameter adjustment is performed on the initial large language model.

[0144] In step S106 of some embodiments, the target large language model is the large language model after the initial large language model has been parameter adjusted. Continuing with the example of step S105, when the target user feedbacks that the answer content is too divergent, the temperature parameter can be lowered. When the feedback information indicates that the answer content is irrelevant to the question, the values ​​of the parameters top_k and top_p can be increased. When the feedback information indicates that the answer content is incoherent, the maximum sampling length parameter can be increased.

[0145] In step S107 of some embodiments, the implementation of this step is basically the same as the implementation of step S103, except that the large language model used is replaced by the target large language model instead of the initial large language model, so it is not described here. The interpretation result output to the target object is the target interpretation text.

[0146] See also Figure 7 The embodiment of the present application also provides a business document interpretation device, which can implement the above-mentioned business document interpretation method, and the device includes:

[0147] An acquisition module 701 is used to acquire a target business file and a question text input by a target object;

[0148] A knowledge base selection module 702 is used to select a target knowledge base from a preset knowledge base according to a target business document;

[0149] The first interpretation module 703 is used to interpret the question text according to the target knowledge base, the target business document and the preset initial large language model to obtain a candidate interpretation text;

[0150] Output module 704, used to output the candidate interpretation text to the target object;

[0151] The parameter determination module 705 is used to determine the target weight parameter according to the feedback information of the target object on the candidate interpretation text;

[0152] A model updating module 706 is used to adjust the parameters of the initial large language model according to the target weight parameter to obtain a target large language model;

[0153] The second interpretation module 707 is used to interpret the question text according to the target knowledge base, the target business file and the target large language model to obtain a target interpreted text.

[0154] The specific implementation of the business document interpretation device is substantially the same as the specific implementation of the business document interpretation method described above, and will not be described in detail herein.

[0155] The embodiment of the present application also provides an electronic device, the electronic device includes a memory and a processor, the memory stores a computer program, and the processor implements the above-mentioned business file interpretation method when executing the computer program. The electronic device can be any intelligent terminal including a tablet computer, a car computer, etc.

[0156] See also Figure 8 , Figure 8 The hardware structure of an electronic device of another embodiment is illustrated, and the electronic device includes:

[0157] The processor 801 may be implemented by a general-purpose CPU (Central Processing Unit), a microprocessor, an application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of the present application;

[0158] The memory 802 can be implemented in the form of a read-only memory (ROM), a static storage device, a dynamic storage device, or a random access memory (RAM). The memory 802 can store an operating system and other application programs. When the technical solution provided in the embodiment of this specification is implemented by software or firmware, the relevant program code is stored in the memory 802, and the processor 801 calls and executes the business file interpretation method of the embodiment of this application;

[0159] Input / output interface 803, used to implement information input and output;

[0160] The communication interface 804 is used to realize the communication interaction between the device and other devices. The communication can be realized through a wired manner (such as USB, network cable, etc.) or a wireless manner (such as mobile network, WI FI, Bluetooth, etc.);

[0161] A bus 805 that transmits information between the various components of the device (e.g., the processor 801, the memory 802, the input / output interface 803, and the communication interface 804);

[0162] The processor 801 , the memory 802 , the input / output interface 803 and the communication interface 804 are connected to each other in communication within the device via a bus 805 .

[0163] An embodiment of the present application also provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, the above-mentioned business file interpretation method is implemented.

[0164] The memory, as a non-transient computer-readable storage medium, can be used to store non-transient software programs and non-transient computer executable programs. In addition, the memory may include a high-speed random access memory, and may also include a non-transient memory, such as at least one disk storage device, a flash memory device, or other non-transient solid-state storage device. In some embodiments, the memory may optionally include a memory remotely disposed relative to the processor, and these remote memories may be connected to the processor via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0165] The business document interpretation method, business document interpretation device, electronic device and storage medium provided by the embodiment of the present application first obtain the target business document and the relevant question text. Then determine the target knowledge base according to the target business document, and then use the initial large language model to combine the target knowledge base and the target business document to perform a preliminary interpretation of the question text to generate a candidate interpretation text. Subsequently, the candidate interpretation text is output to the target object, and the feedback information of the target object on the candidate interpretation text is obtained. Then adjust the target weight parameter according to the feedback information so that the model can be dynamically optimized according to actual needs, thereby improving the relevance and accuracy of the file interpretation. Then optimize the initial large language model to the target large language model through parameter adjustment, ensuring that the model can meet the needs of the target object. Finally, use the optimized target large language model in combination with the target knowledge base and the target business document to interpret the question text again, and generate a high-quality target interpretation text. The method of this embodiment optimizes the parameters of the large language model that generates the interpretation text by screening the target knowledge bureau that matches the target business document and promptly responding to the feedback information of the target object, so as to obtain more accurate interpretation results, thereby improving the accuracy of document interpretation, and ultimately providing the business personnel of the financial institution with a more efficient, accurate and reliable document interpretation method, helping them to quickly grasp industry changes and optimize business development strategies.

[0166] The embodiments described in the embodiments of the present application are intended to more clearly illustrate the technical solutions of the embodiments of the present application and do not constitute a limitation on the technical solutions provided in the embodiments of the present application. Those skilled in the art will appreciate that with the evolution of technology and the emergence of new application scenarios, the technical solutions provided in the embodiments of the present application are also applicable to similar technical problems.

[0167] Those skilled in the art will appreciate that the technical solutions shown in the figures do not constitute a limitation on the embodiments of the present application, and may include more or fewer steps than shown in the figures, or a combination of certain steps, or different steps.

[0168] The device embodiments described above are merely illustrative, and the units described as separate components may or may not be physically separated, that is, they may be located in one place or distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0169] Those skilled in the art will appreciate that all or some of the steps in the methods disclosed above, and the functional modules / units in the systems and devices may be implemented as software, firmware, hardware, or a suitable combination thereof.

[0170] The terms "first", "second", "third", "fourth", etc. (if any) in the specification of the present application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0171] It should be understood that in the present application, "at least one (item)" means one or more, and "plurality" means two or more. "And / or" is used to describe the association relationship of associated objects, indicating that three relationships may exist. For example, "A and / or B" can mean: only A exists, only B exists, and A and B exist at the same time, where A and B can be singular or plural. The character " / " generally indicates that the objects associated before and after are in an "or" relationship. "At least one of the following" or similar expressions refers to any combination of these items, including any combination of single or plural items. For example, at least one of a, b or c can mean: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, c can be single or multiple.

[0172] In the several embodiments provided in the present application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of the above units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0173] The units described above as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0174] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0175] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including multiple instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of various embodiments of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (Read-Only Memory, referred to as ROM), random access memory (Random Access Memory, referred to as RAM), disk or optical disk and other media that can store programs.

[0176] The preferred embodiments of the present invention are described above with reference to the accompanying drawings, but the scope of the rights of the present invention is not limited thereto. Any modification, equivalent substitution and improvement made by a person skilled in the art without departing from the scope and essence of the present invention should be within the scope of the rights of the present invention.

Claims

1. A method for interpreting a business document, characterized in that: The method comprises: Obtain the target business file and question text input by the target object; Selecting a target knowledge base from a preset knowledge base according to the target business document; Interpreting the question text according to the target knowledge base, the target business document and a preset initial large language model to obtain a candidate interpretation text; Outputting the candidate interpretation text to the target object; Determining a target weight parameter according to feedback information of the target object on the candidate interpretation text; Adjusting the parameters of the initial large language model according to the target weight parameter to obtain a target large language model; The question text is interpreted according to the target knowledge base, the target business file and the target large language model to obtain a target interpreted text.

2. The method according to claim 1, characterized in that The step of interpreting the question text according to the target knowledge base, the target business file and the preset initial large language model to obtain a candidate interpretation text includes: Performing feature vectorization on the target knowledge base to obtain a knowledge base vector; Performing feature vectorization on the target business file to obtain a target file vector; Performing attention feature fusion on the knowledge base vector and the target file vector to obtain a reference attention fusion vector; Performing feature vectorization on the question text to obtain a question vector; The reference attention fusion vector and the question vector are subjected to text generation through the initial large language model to obtain the candidate interpretation text.

3. The method according to claim 2, characterized in that The step of performing feature vectorization on the target business file to obtain a target file vector includes: Extracting the target business file to obtain at least two business sub-files; wherein any two adjacent business sub-files overlap; Performing feature vectorization on each of the business sub-files to obtain a local vector; Feature fusion is performed on each of the local vectors to obtain the target file vector.

4. The method according to claim 3, characterized in that The feature vectorization of each of the business sub-files to obtain a local vector includes: Extracting text from the business sub-file to obtain a target text; Extracting position features of the target text to obtain a text position vector; Performing text semantic extraction on the business sub-file to obtain a text semantic vector; Perform feature fusion according to the text position vector and the text semantic vector to obtain a text fusion feature vector; Extracting images from the business sub-file to obtain a target image; Extracting position features of the target image to obtain an image position vector; Performing image description processing on the target image to obtain an image meaning vector; Perform feature fusion according to the image position vector and the image meaning vector to obtain an image fusion feature vector; A self-attention calculation is performed on the text fusion feature vector according to the image fusion feature vector to obtain the local vector.

5. The method according to any one of claims 1 to 4, characterized in that: The step of outputting the candidate interpretation text to the target object includes: Performing authenticity evaluation on the candidate interpretation text according to the target knowledge base and the target business document to obtain an authenticity score; Performing a correlation evaluation on the question text and the candidate interpretation text to obtain a correlation degree score; The candidate interpretation text is output to the target object according to the relevance score and the authenticity score.

6. The method according to claim 5, characterized in that The step of outputting the candidate interpretation text to the target object according to the association degree score and the authenticity score includes: Integrate and calculate the relevance score and the authenticity score to obtain a target score; If the target score is greater than or equal to a preset benchmark threshold, the candidate interpretation text is output to the target object.

7. The method according to claim 6, characterized in that If the target score is greater than or equal to a preset reference threshold, before outputting the candidate interpretation text to the target object, the method further includes: Performing compliance assessment on the candidate interpretation text to obtain a compliance score; Performing weighted processing on the compliance score and the relevance score to obtain a sub-score; The authenticity score and the sub-score are numerically multiplied to obtain a reference score; wherein the authenticity score is 0 or 1; The target score is updated according to the reference score.

8. A business document interpretation device, characterized in that: The device comprises: An acquisition module is used to acquire a target business file and a question text input by a target object; A knowledge base selection module, used to select a target knowledge base from a preset knowledge base according to the target business document; A first interpretation module is used to interpret the question text according to the target knowledge base, the target business document and a preset initial large language model to obtain a candidate interpretation text; An output module, used for outputting the candidate interpretation text to the target object; A parameter determination module, used to determine a target weight parameter according to feedback information of the target object on the candidate interpretation text; A model updating module, used for adjusting the parameters of the initial large language model according to the target weight parameter to obtain a target large language model; The second interpretation module is used to interpret the question text according to the target knowledge base, the target business file and the target large language model to obtain a target interpreted text.

9. An electronic device, characterized in that: The electronic device comprises a memory and a processor, the memory stores a computer program, and the processor implements the method according to any one of claims 1 to 7 when executing the computer program.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images