Property clue mining method and device, computer equipment and storage medium

Abstract

The invention discloses a property clue mining method and device, computer equipment and a storage medium, and belongs to the technical field of artificial intelligence. A text processing model and a clue classification model are integrated in a property clue mining platform, firstly, a case text is received, summary information of the case text is extracted, a word embedding matrix of the text processing model is used for converting a summary into a semantic feature vector, and then a text feature sequence is constructed based on the semantic feature vector; and carrying out local feature extraction on the sequence by applying a convolution kernel of the clue classification model, fusing the local features and semantic features to form a comprehensive feature fusion vector, and finally, accurately matching and outputting a target property clue tag related to a case based on the fusion vector. The invention further relates to the technical field of block chains, and the case text is stored on a block chain node. By reducing the dependence of artificial feature engineering, the operation difficulty and cost are reduced, and the precision and efficiency of property clue mining are improved.

Description

Technical Field

[0001] The present application belongs to the field of artificial intelligence technology, and specifically relates to a property clue mining method, device, computer equipment and storage medium. Background Art

[0002] In the financial field, non-performing assets refer to debt assets that have fallen into a state of default or have been confirmed to be unpayable. Such assets are difficult to directly convert into cash flow or have extremely low cash value, which often brings heavy burdens to financial institutions and creditors. In order to reduce this burden and safeguard the rights and interests of creditors, it is particularly important to effectively collect non-performing assets. In this process, the core task is to dig deep into the debtor's property clues, which are the key to recovering the debt and reducing losses.

[0003] At present, when dealing with non-performing assets, financial institutions mainly rely on three strategies: self-collection, litigation and entrusting third parties to collect debts. Whether the strategies can be effectively implemented depends on whether the debtor's executable property clues can be successfully discovered and confirmed. Therefore, facing corporate debtors with operating difficulties or even on the verge of bankruptcy, finding and confirming effective executable property clues has become the top priority of recovery work.

[0004] On the technical level, although traditional machine learning technologies such as support vector machines and neural networks have been applied to the mining of property clues and have achieved certain results, their limitations are becoming increasingly apparent. These methods are highly dependent on complex artificial feature engineering, which not only increases the difficulty and cost of operation, but also makes the performance of the model susceptible to data quality and sample size limitations.

[0005] Therefore, exploring more efficient, intelligent and adaptable technical means to optimize the property clue mining process has become an urgent issue to be solved in the field of non-performing asset recovery. Summary of the invention

[0006] The purpose of the embodiments of the present application is to propose a property clue mining method, apparatus, computer equipment and storage medium to solve the technical problems of existing property clue mining solutions, which are highly dependent on complex manual feature engineering, increase the difficulty and cost of operation, and make the performance of the model susceptible to limitations of data quality and sample size.

[0007] In order to solve the above technical problems, the present application embodiment provides a property clue mining method, which adopts the following technical solution:

[0008] A property clue mining method is applied to a property clue mining platform. The property clue mining platform includes a text processing model and a clue classification model. The property clue mining method includes:

[0009] Receive a property clue mining instruction, obtain the case text of the pending property case, and obtain the pending case text;

[0010] Import the case text to be processed into the property clue mining platform, extract summary information of the case text to be processed, and obtain summary text;

[0011] Use the word embedding matrix in the text processing model to semantically encode the summary text and obtain a semantic feature vector;

[0012] Construct a text feature sequence based on the semantic feature vector, and use the convolution kernel in the clue classification model to perform a local convolution operation on the text feature sequence to obtain a local feature vector;

[0013] Concatenate the semantic feature vector and the local feature vector to obtain a feature fusion vector;

[0014] A target property clue label matching the property case to be processed is determined based on the feature fusion vector, and the target property clue label is output.

[0015] Furthermore, before the step of using the word embedding matrix in the text processing model to semantically encode the summary text to obtain the semantic feature vector, the method further includes:

[0016] Obtaining a first historical case text from a preset case database;

[0017] Perform keyword extraction on the first historical case text, and select legal segmentation text from the keyword extraction results of the first historical case text;

[0018] Import the legal segmentation text into the original vocabulary of the preset text processing model;

[0019] The original vocabulary is vectorized to obtain a word embedding matrix.

[0020] Furthermore, after the step of vectorizing the original word list to obtain the word embedding matrix, the method further includes:

[0021] Acquire the second historical case text from a preset case database, and perform data annotation on the second historical case text;

[0022] Constructing training data set and test data set based on the annotated second historical case text;

[0023] Pre-train the preset text processing model using the training dataset;

[0024] Use the test dataset to test the model performance of the pre-trained preset text processing model.

[0025] Furthermore, the text processing model is a pre-trained language model based on the Transformer architecture. The Transformer architecture includes several encoding layer structures and several decoding layer structures. The steps of pre-training the preset text processing model using the training data set specifically include:

[0026] The historical case texts in the training dataset are mapped into vector representations through word embedding matrices, and a training sample feature matrix is ​​constructed;

[0027] Import the training sample feature matrix into several encoding layers of the Transformer architecture in sequence according to the feature hierarchy relationship to perform encoding operations on the training sample features;

[0028] Obtain the feature encoding result output by the encoding layer, and input the feature encoding result into the decoding layer to perform a decoding operation on the encoded feature;

[0029] Obtaining a feature decoding result, and comparing the feature decoding result with a first true label of the historical case text to obtain a first prediction error, wherein the first true label is a marked text semantic label;

[0030] The parameters of the Transformer architecture are iteratively updated based on the first prediction error until the model is fitted to obtain a preset text processing model.

[0031] Furthermore, the clue classification model is a pre-trained neural network model based on a text convolutional neural network architecture, and a plurality of convolution kernels are preset in the text convolutional neural network. Before the step of constructing a text feature sequence based on a semantic feature vector and performing a local convolution operation on the text feature sequence using the convolution kernel in the clue classification model to obtain a local feature vector, the method further includes:

[0032] Parse the feature encoding result output by the encoding layer to obtain the feature encoding vector output by the encoding layer;

[0033] Constructing a feature coding sequence based on the feature coding vector;

[0034] Determine an adapted target convolution kernel among several convolution kernels in a preset clue classification model;

[0035] Use the target convolution kernel to perform a local convolution operation on the feature coding sequence to obtain a local coding vector; concatenate the feature coding vector and the local coding vector to obtain a coding fusion vector;

[0036] Determine the property clue label that matches the historical case text based on the feature fusion vector to obtain the historical property clue label;

[0037] Comparing the historical property clue label with a second true label of the historical case text to obtain a second prediction error, wherein the second true label is the annotated property clue label;

[0038] The parameters of the clue classification model are iteratively updated based on the second prediction error until the model is fitted to obtain a preset text processing model.

[0039] Furthermore, the step of constructing a feature coding sequence based on the feature coding vector specifically includes:

[0040] Screening the feature coding vector to obtain a high-order feature coding vector;

[0041] Construct feature encoding sequences based on high-order feature encoding vectors.

[0042] Further, the parameters of the clue classification model are iteratively updated based on the second prediction error.

[0043] After the model is fitted and the preset text processing model is obtained, the following steps are also included:

[0044] Configure initial model weight parameters for the text processing model and clue classification model;

[0045] A cross entropy loss function is set, and based on the first prediction error and the second prediction error, the initial model weight parameters are dynamically updated with minimization of the cross entropy loss function as the optimization goal to obtain the optimal model weight parameters;

[0046] Configure the text processing model and clue classification model in the property clue mining platform based on the optimal model weight parameters.

[0047] In order to solve the above technical problems, the embodiment of the present application further provides a property clue mining device, which adopts the following technical solution:

[0048] A property clue mining device, the property clue mining device runs on a property clue mining platform, the property clue mining platform includes a text processing model and a clue classification model, the property clue mining device includes:

[0049] A text extraction module, used to receive a property clue mining instruction, obtain the case text of the pending property case, and obtain the pending case text;

[0050] A summary extraction module is used to import the case text to be processed into the property clue mining platform, extract summary information of the case text to be processed, and obtain summary text;

[0051] A semantic encoding module is used to semantically encode the summary text using the word embedding matrix in the text processing model to obtain a semantic feature vector;

[0052] The local convolution module is used to construct a text feature sequence based on the semantic feature vector, and use the convolution kernel in the clue classification model to perform a local convolution operation on the text feature sequence to obtain a local feature vector;

[0053] A vector concatenation module is used to concatenate the semantic feature vector and the local feature vector to obtain a feature fusion vector;

[0054] The clue matching module is used to determine the target property clue label that matches the property case to be processed based on the feature fusion vector, and output the target property clue label.

[0055] 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:

[0056] A computer device comprises a memory and a processor, wherein the memory stores computer-readable instructions, and the processor implements the steps of any one of the above-mentioned property clue mining methods when executing the computer-readable instructions.

[0057] 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:

[0058] A computer-readable storage medium stores computer-readable instructions, which, when executed by a processor, implement the steps of any one of the above-mentioned property clue mining methods.

[0059] Compared with the prior art, the embodiments of the present application have the following beneficial effects:

[0060] The present application discloses a property clue mining method, device, computer equipment and storage medium, which belongs to the field of artificial intelligence technology. By integrating the two models of text processing and clue classification in the property clue mining platform, the case text is first received, its summary information is extracted, and the word embedding matrix of the text processing model is used to convert the summary into a semantic feature vector, and then a text feature sequence is constructed based on the semantic feature vector, and the convolution kernel sequence of the clue classification model is applied to extract local features, and the local features are fused with the semantic features to form a comprehensive feature fusion vector, and finally the target property clue label related to the case is accurately matched and output based on the fusion vector. The present application automatically extracts the case text summary, generates semantic features using the word embedding matrix, and captures local features in combination with the convolution operation, which effectively solves the dependence of traditional property clue mining on artificial feature engineering and reduces the difficulty and cost of operation. At the same time, the feature fusion strategy enhances the generalization ability of the model, reduces the dependence on data quality and sample size, and improves the accuracy and efficiency of property clue mining. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] 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.

[0062] Figure 1 An exemplary system architecture diagram is shown in which the present application can be applied;

[0063] Figure 2 A flow chart showing an embodiment of a property clue mining method according to the present application is shown;

[0064] Figure 3 A schematic structural diagram of an embodiment of a property clue mining device according to the present application is shown;

[0065] Figure 4 A schematic structural diagram of an embodiment of a computer device according to the present application is shown. DETAILED DESCRIPTION

[0066] 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.

[0067] 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.

[0068] 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.

[0069] like Figure 1As shown, the system architecture 100 may include terminal devices 101, 102, 103, a network 104 and a server 105. The network 104 is used to provide a medium for communication links between the terminal devices 101, 102, 103 and the server 105. The network 104 may include various connection types, such as wired, wireless communication links or optical fiber cables, etc.

[0070] Users can use terminal devices 101, 102, 103 to interact with server 105 through network 104 to receive or send messages, etc. Various communication client applications can be installed on terminal devices 101, 102, 103, such as web browser applications, shopping applications, search applications, instant messaging tools, email clients, social platform software, etc.

[0071] Terminal devices 101, 102, 103 can be various electronic devices with display screens and supporting web browsing, including but not limited to smart phones, tablet computers, e-book readers, MP3 players (Moving Picture Experts Group Audio Layer III), MP4 (Moving Picture Experts Group Audio Layer IV) players, laptop computers and desktop computers, etc.

[0072] Server 105 can be a server that provides various services, such as a background server that provides support for pages displayed on terminal devices 101, 102, and 103. The server can be an independent server or 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, content distribution networks (CDNs), and big data and artificial intelligence platforms.

[0073] It should be noted that the property clue mining method provided in the embodiment of the present application is generally executed by a server, and accordingly, the property clue mining device is generally set in the server.

[0074] 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.

[0075] Continue to refer Figure 2, showing a flow chart of an embodiment of the property clue mining method according to the present application. The present application embodiment 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.

[0076] 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.

[0077] In this embodiment, the application discloses a property clue mining method, which is applied to a property clue mining platform. The property clue mining platform includes a text processing model and a clue classification model. The two work together to more effectively automatically extract features that are useful for property clue classification from text data. By automatically learning the deep semantic representation and local features of the text, it is expected to significantly reduce the reliance on artificial feature engineering and alleviate the model's limitations on data quality and sample size.

[0078] In a specific embodiment of the present application, the text processing model is a pre-trained language model based on the Transformer architecture, taking the Bert (Bidirectional Encoder Representations from Transformers) model as an example. A major advantage of the Bert model is that it can automatically learn the deep semantic representation of text without too much manual feature engineering. The clue classification model is a pre-trained neural network model based on the text convolutional neural network architecture, taking the TextCNN (Text Convolutional Neural Networks) model as an example. The local convolution operation of the TextCNN model can capture local patterns in the text, and these local patterns can improve the recognition accuracy of property clues.

[0079] By combining the Bert model with the TextCNN model, we can use Bert's powerful semantic understanding ability to capture the contextual information of the text, while taking advantage of TextCNN's advantages in local feature extraction. The two work together to more effectively automatically extract features that are useful for property clue classification from text data, thereby greatly reducing dependence on manual feature engineering.

[0080] The property clue mining platform of this application is built based on the Bert+TextCNN model. The integrated non-performing asset mining text classification method significantly reduces the dependence on manual feature engineering by automatically learning the deep semantic representation and local features of the text, and alleviates the model's limitations on data quality and sample size.

[0081] In this embodiment, continue to refer to Figure 2 The property clue mining method disclosed in this application specifically includes the following steps:

[0082] S201, receiving a property clue mining instruction, obtaining the case text of the pending property case, and obtaining the pending case text.

[0083] Specifically, by receiving property clue mining instructions, obtaining the case text of pending property cases, and obtaining the pending case text, the system realizes efficient processing of specific property case texts. In practical applications, when a user needs to mine potential clues for a property case, the system will receive the corresponding mining instructions, and then the system will retrieve or obtain the detailed text information of the case from the database in real time as the pending case text. For example, in a case involving asset transfer, the system can quickly capture text content such as case descriptions, relevant evidence documents, and witness testimonies.

[0084] S202, importing the case text to be processed into the property clue mining platform, extracting summary information of the case text to be processed, and obtaining a summary text.

[0085] Specifically, by importing the text of pending cases into the property clue mining platform, the platform's text processing capabilities are used to automatically extract case summary information. The case summary is a high-level summary of the case content, usually containing key information of the case, such as case type, persons involved, main facts, focus of dispute, and judgment results. This not only improves processing efficiency, but also makes it easier for users to quickly understand the core content of the case. For example, when handling a complex financial dispute case, the system can automatically extract key information from the lengthy case text, such as case background, focus of dispute, amount involved, etc., and generate a concise and clear summary text.

[0086] The property clue mining platform also includes a text processing module. After uploading the case text to the platform, the platform's built-in text processing module will be activated. This module integrates natural language processing (NLP) technology and can intelligently identify key information in the text, such as time, place, people, events, etc. By extracting and reorganizing these elements and combining them with the preset summary generation algorithm, the platform can automatically generate concise and accurate case summary information, which not only improves the efficiency of information processing, but also ensures the accuracy and comprehensiveness of the summary.

[0087] S203, using the word embedding matrix in the text processing model to semantically encode the summary text to obtain a semantic feature vector.

[0088] Specifically, by applying the word embedding matrix in the text processing model, the summary text is deeply vectorized and semantically encoded, thereby converting it into a feature vector rich in semantic information. As a key technology in natural language processing, the word embedding matrix can map the words in the text to a high-dimensional space, so that similar words are close in space, and then capture the semantic relationship between words. In the context of property clue mining, semantic encoding can help the model better understand the semantic content in the case summary, such as the relationship between the people involved in the case, the path of property transfer, etc. For example, when dealing with an economic crime case, the model can identify key words such as "illegal fundraising" and "capital flow", and convert them into corresponding semantic feature vectors through the word embedding matrix.

[0089] S204, constructing a text feature sequence based on the semantic feature vector, and using the convolution kernel in the clue classification model to perform a local convolution operation on the text feature sequence to obtain a local feature vector.

[0090] Specifically, by constructing text feature sequences and using the convolution kernels in the clue classification model to perform local convolution operations on these sequences, effective extraction of local features of the text is achieved. The text feature sequence is constructed based on the semantic feature vector, which not only retains the original information of the text, but also incorporates rich semantic connotations. As a tool for feature extraction, the convolution kernel can automatically learn and capture key features in local areas, such as keyword combinations, sentence structures, etc. For example, when dealing with a property dispute case, the model may identify key evidence mentions such as "real estate transfer contracts" and "bank transfer records" through convolution operations, and then generate corresponding local feature vectors. These vectors not only reflect the specific content of the text, but also imply deep information related to property clues.

[0091] S205, concatenating the semantic feature vector and the local feature vector to obtain a feature fusion vector.

[0092] Specifically, the strategy of concatenating semantic feature vectors with local feature vectors to form feature fusion vectors aims to comprehensively integrate the multi-level information of the text. The semantic feature vector captures the global semantic information of the text, while the local feature vector focuses on the local details and key features of the text. Through feature vector concatenation, the feature fusion vector retains the overall meaning of the text while highlighting important local features. For example, when dealing with a complex financial case, the feature fusion vector may contain both the overall background information of the case and highlight key details such as the flow of funds and transaction patterns between the people involved. The fusion of this information helps to more accurately identify potential property clues.

[0093] S206, determining a target property clue label that matches the property case to be processed based on the feature fusion vector, and outputting the target property clue label.

[0094] Specifically, the feature fusion vector is used to accurately determine and output the target property clue label of the property case to be processed. The feature fusion vector integrates multiple aspects of information such as the semantics and local features of the text, providing a comprehensive and in-depth understanding basis for the model. In the process of determining the target property clue label based on the feature fusion vector, the feature fusion vector needs to be passed as input to the fully connected layer of the clue classification model. In the fully connected layer, each neuron is connected to all neurons in the previous layer (i.e., the feature fusion layer), and more abstract and advanced feature representations are extracted through weighted summation and nonlinear activation functions (such as ReLU or Sigmaid). Subsequently, these feature representations are further passed to the neurons in the classification layer, and each neuron corresponds to a predefined target property clue label. By calculating the output values ​​of these neurons and applying the softmax function for normalization, the probability distribution of each label is obtained. Finally, the label with the highest probability is selected as the target property clue label matching the property case to be processed for output.

[0095] After obtaining the debtor or guarantor information accurately identified by the target property clue label, the system can more systematically evaluate the property clue status, actual performance ability, and capital flow behavior (such as transfer-in and transfer-out transactions) of these objects, and make a comprehensive judgment based on the key information of the case (such as the cause of the case, the amount involved, the identity of the plaintiff and the defendant, the trial results and time, etc.). This process helps the system to accurately stratify the difficulty of debt repayment for creditors or guarantors, so as to facilitate the subsequent formulation of differentiated collection strategies. This refined management and efficient operation mode also significantly improves work efficiency and promotes the optimization and upgrading of the entire debt management process.

[0096] In the above embodiment, this application automatically extracts case text summaries, generates semantic features using word embedding matrices, and captures local features in combination with convolution operations, effectively solving the reliance of traditional property clue mining on manual feature engineering and reducing operational difficulty and cost. At the same time, the feature fusion strategy enhances the generalization ability of the model, reduces reliance on data quality and sample size, and improves the accuracy and efficiency of property clue mining.

[0097] Furthermore, before the step of using the word embedding matrix in the text processing model to semantically encode the summary text to obtain the semantic feature vector, the method further includes:

[0098] Obtaining a first historical case text from a preset case database;

[0099] Perform keyword extraction on the first historical case text, and select legal segmentation text from the keyword extraction results of the first historical case text;

[0100] Import the legal segmentation text into the original vocabulary of the preset text processing model;

[0101] The original vocabulary is vectorized to obtain a word embedding matrix.

[0102] In this embodiment, the text in the legal field is converted into a deep semantic feature vector. First, a representative first historical case text is selected from the case database. Subsequently, keyword extraction technology is used to extract legal related vocabulary, that is, legal segmentation text, from a large number of legal texts such as legal documents, precedents, and legal provisions, to identify and extract professional vocabulary and terms unique to the legal field. These legal segmentation texts include legal terms, legal concepts, names of legal institutions, legal file names, etc. The legal segmentation text extraction effectively eliminates non-legal related interference items and improves the accuracy of the analysis. Next, the screened legal segmentation text is integrated into the original vocabulary of the text processing model, injecting professional legal context knowledge into the text processing model. Finally, by vectorizing the original vocabulary, that is, constructing a word embedding matrix, the transformation of the text from the literal to the semantic level is realized.

[0103] For example, when training, the BERT model converts each word or token in the text into a fixed-size vector representation based on a predefined vocabulary (also called a word embedding matrix or embedding layer). By adding legal domain-specific vocabulary to the original vocabulary of the BERT model, it ensures that these words can be better represented when the model processes legal texts.

[0104] Through the above steps, the performance of the model in legal text processing is improved by customizing the original vocabulary of the text processing model and by expanding its vocabulary to include legal domain-specific vocabulary.

[0105] Furthermore, after the step of vectorizing the original word list to obtain the word embedding matrix, the method further includes:

[0106] Acquire the second historical case text from a preset case database, and perform data annotation on the second historical case text;

[0107] Constructing training data set and test data set based on the annotated second historical case text;

[0108] Pre-train the preset text processing model using the training dataset;

[0109] Use the test dataset to test the model performance of the pre-trained preset text processing model.

[0110] In this embodiment, first, the second historical case text is extracted from a rich preset case database. The second historical case text can be the same as the first historical case text or different from the first historical case text. Then, data annotation is performed on the second historical case text, and text semantic labels and property clue labels are annotated. Subsequently, a training data set and a test data set are constructed. The training data set is used to perform deep pre-training on the preset text processing model, aiming to enable the model to fully learn the characteristics and rules of the case text. Finally, the pre-trained model is performance tested using the test data set to verify its accuracy, efficiency and generalization ability when processing similar case texts.

[0111] Through the above steps, by extracting and annotating data from historical case texts, we built high-quality training and test data sets, effectively improving the pre-training effect of the text processing model. After the text processing model learns the case text features on the training data set, it is verified through the test data set to test the model performance.

[0112] Furthermore, the text processing model is a pre-trained language model based on the Transformer architecture. The Transformer architecture includes several encoding layer structures and several decoding layer structures. The steps of pre-training the preset text processing model using the training data set specifically include:

[0113] The historical case texts in the training dataset are mapped into vector representations through word embedding matrices, and a training sample feature matrix is ​​constructed;

[0114] Import the training sample feature matrix into several encoding layers of the Transformer architecture in sequence according to the feature hierarchy relationship to perform encoding operations on the training sample features;

[0115] Get the feature encoding result output by the encoding layer and input the feature encoding result into the decoding layer.

[0116] To perform a decoding operation of the encoded feature;

[0117] Obtaining a feature decoding result, and comparing the feature decoding result with a first true label of the historical case text to obtain a first prediction error, wherein the first true label is a marked text semantic label;

[0118] The parameters of the Transformer architecture are iteratively updated based on the first prediction error until the model is fitted to obtain a preset text processing model.

[0119] In this embodiment, a pre-trained language model based on the Transformer architecture is used to deeply optimize the processing of historical case texts. The Transformer architecture, with its unique encoding-decoding layer structure, can efficiently capture complex semantic relationships and contextual dependencies in the text. In the pre-training process, the historical case texts in the training data set are first converted into high-dimensional vector representations through word embedding technology, which effectively retains the semantic information of the text and constructs a feature matrix of the training sample. Subsequently, the feature matrix is ​​progressively encoded using the multi-layer encoding layer of the Transformer architecture, and each layer of the encoding layer further refines and abstracts the text features, thereby enhancing the model's ability to understand the deep semantics of the case text. After the encoding is completed, the decoding layer is responsible for reversely converting the encoded features back to an interpretable text form, and by comparing the decoding results with the true semantic labels (i.e., the first true labels), the prediction error is calculated, and the model parameters are iteratively updated accordingly, which not only improves the accuracy of the model's processing of the case text, but also ensures that the model can stably and efficiently fit the training data through continuous parameter optimization, and finally obtains a preset text processing model that can extract text semantic features.

[0120] Take the BERT model as an example of a text processing model. In BERT, the encoder layer refers to the encoder part of the Transformer model. The Transformer model consists of encoders and decoders. These encoder layers are stacked to capture different levels of information in text data. Each encoder layer contains a self-attention mechanism.

[0121] (se l f-attent i on mechanism) and feed-forward neural network, which work together to understand and transform the input text data.

[0122] When encoding sample features, each encoder layer outputs a feature encoding vector. Each feature encoding vector is specially marked with CLS. CLS (Classification) is a special marker (i.e., word fragment token) added to the beginning of the input sequence when the BERT model processes text data. This special marker is usually expressed as [CLS] (in some implementations it may be [CLS], in other implementations it may be [SEP] or other). When processing the input, the BERT model will embed this special marker.

[0123] The word embedding is fed into the model together with the embeddings of other words in the input text sequence.

[0124] After processing by multiple encoder layers of BERT, the output of the [CLS] token of each encoder layer (i.e., its hidden state) can be regarded as the "sentence vector" or "sequence representation" of the entire input sequence. This vector contains information about the entire input sequence and can be used for various downstream tasks such as text classification, sentiment analysis, etc.

[0125] In a specific embodiment of the present application, in the BERT model, in addition to the first input layer, 12 encoder layers are also included, and the vectors of the first tag (CLS) of each encoder layer carry sentence semantic information. As the number of encoder layers increases, these vectors gradually transition from capturing low-level semantic details to higher-level abstract semantic features. In order to fully capture lexical features and their semantic expressions at different levels, the feature encoding vectors carrying CLS output by the last four encoder layers of BERT can be used as the input of TextCNN. TextCNN can further extract and fuse these multi-level semantic information through its convolution and pooling operations, and ultimately use them for classification tasks.

[0126] Through the above steps, the pre-trained language model of the Transformer architecture is used to deeply explore the semantic features of historical case texts through word embedding, multi-layer encoding and decoding. By continuously iterating and optimizing model parameters, the prediction error is significantly reduced, and efficient and accurate processing of case texts is achieved, which significantly improves the performance of the text processing model.

[0127] Furthermore, the clue classification model is a pre-trained neural network model based on a text convolutional neural network architecture, and a plurality of convolution kernels are preset in the text convolutional neural network. Before the step of constructing a text feature sequence based on a semantic feature vector and performing a local convolution operation on the text feature sequence using the convolution kernel in the clue classification model to obtain a local feature vector, the method further includes:

[0128] Parse the feature encoding result output by the encoding layer to obtain the feature encoding vector output by the encoding layer;

[0129] Constructing a feature coding sequence based on the feature coding vector;

[0130] Determine an adapted target convolution kernel among several convolution kernels in a preset clue classification model;

[0131] Use the target convolution kernel to perform a local convolution operation on the feature coding sequence to obtain a local coding vector; concatenate the feature coding vector and the local coding vector to obtain a coding fusion vector;

[0132] Determine the property clue label that matches the historical case text based on the feature fusion vector to obtain the historical property clue label;

[0133] Comparing the historical property clue label with a second true label of the historical case text to obtain a second prediction error, wherein the second true label is the annotated property clue label;

[0134] The parameters of the clue classification model are iteratively updated based on the second prediction error until the model is fitted to obtain a preset text processing model.

[0135] In this embodiment, the clue classification model is a pre-trained neural network model based on the text convolutional neural network architecture. Taking the TextCNN model as an example, a number of convolution kernels are preset in the text convolutional neural network. First, by parsing the feature coding vector output by the coding layer, a feature coding sequence that can reflect the deep semantics of the text is constructed. Subsequently, in the preset clue classification model (i.e., the TextCNN model), a target convolution kernel adapted to the feature coding vector specification is selected, and several convolution kernels of different specifications in the preset clue classification model, such as 2*4, 3*4, and 4*4 convolution kernels, are preset. The feature coding sequence is subjected to a local convolution operation using the adapted target convolution kernel to extract a local feature coding vector, i.e., an n-gram feature vector. n-gram refers to a combination of n items that appear continuously in the text, where "n" is a positive integer, indicating the number of items in the combination. The n-gram feature can capture local semantic information in the text, including keywords, phrases, and their combination methods, which helps to understand the overall meaning of the text and perform tasks such as text classification. The n-gram feature vector not only retains the global features, but also strengthens the local details and improves the richness of the feature representation. Next, the global feature encoding vector and the local encoding vector are cleverly spliced ​​together to form an encoding fusion vector, which enhances the model's ability to identify property clues. Based on this fusion vector, the model can accurately match and determine the property clue labels in the historical case text, and calculate the prediction error by comparing it with the true label. Finally, the parameters of the clue classification model are iteratively updated to continuously optimize the model performance until the best fit is achieved.

[0136] Through the above steps, the combination of TextCNN architecture and pre-trained model effectively improves the accuracy of property clue classification. By using multi-layer feature encoding and local convolution operations, the model's understanding of the deep semantics of case text and feature extraction capabilities are enhanced. Through feature fusion and iterative optimization, the prediction error is reduced and accurate matching of property clue labels is achieved.

[0137] Furthermore, the step of constructing a feature coding sequence based on the feature coding vector specifically includes:

[0138] Screening the feature coding vector to obtain a high-order feature coding vector;

[0139] Construct feature encoding sequences based on high-order feature encoding vectors.

[0140] In this embodiment, when constructing a feature coding sequence, the feature coding vectors are firstly screened to extract high-order feature coding vectors, which often contain deeper semantic information. The screening process is to focus on those feature coding vectors with high-level semantics output by the deep encoding layer. Through such processing, the constructed feature coding sequence can not only retain the basic semantics of the text, but also highlight its deep meaning in a complex context.

[0141] Further, the parameters of the clue classification model are iteratively updated based on the second prediction error.

[0142] After the model is fitted and the preset text processing model is obtained, the following steps are also included:

[0143] Configure initial model weight parameters for the text processing model and clue classification model;

[0144] A cross entropy loss function is set, and based on the first prediction error and the second prediction error, the initial model weight parameters are dynamically updated with minimization of the cross entropy loss function as the optimization goal to obtain the optimal model weight parameters;

[0145] Configure the text processing model and clue classification model in the property clue mining platform based on the optimal model weight parameters.

[0146] In this embodiment, by configuring the initial model weight parameters and using the cross entropy loss function as the optimization criterion, these parameters are dynamically adjusted to minimize the prediction error, which not only ensures the efficiency of the text processing model and the clue classification model during the training process, but also obtains the optimal model weight parameters through iterative optimization, thereby significantly improving the accuracy and performance of the model in the property clue mining task. Finally, the optimal model weight parameters are deployed on the property clue mining platform.

[0147] During the training of the text processing model and clue classification model, the cross entropy loss function is used to measure the difference between the model's predicted probability distribution and the true label probability distribution. When setting up specifically, first define the loss function as the cross entropy value between the probability distribution predicted by the model and the true label. Then, based on the first prediction error and the second prediction error, the gradient of the loss function with respect to the model weight parameters is calculated by the back propagation algorithm. With minimizing the cross entropy loss function as the optimization goal, the initial model weight parameters are dynamically updated using the gradient descent method or a more advanced optimization algorithm (such as Adam, RMSprop, etc.). Through multiple iterative trainings, the weight parameters are gradually adjusted until the loss function value converges to the minimum. At this time, the weight parameters obtained are the optimal model weight parameters.

[0148] In this embodiment, the electronic device (eg Figure 1 The server shown in the figure) can receive instructions or obtain 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 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.

[0149] It should be emphasized that in order to further ensure the privacy and security of the above case texts, the above case texts can also be stored in a blockchain node.

[0150] 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.

[0151] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing related hardware through computer-readable instructions, and the computer-readable instructions can be stored in a computer-readable storage medium. When the computer-readable instructions are executed, they 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).

[0152] 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.

[0153] Further references Figure 3 , as a response to the above Figure 2The present application provides an embodiment of a property clue mining device, and the device embodiment is Figure 2 Corresponding to the method embodiment shown, the device can be specifically applied to various electronic devices.

[0154] like Figure 3 As shown, the property clue mining device 300 described in this embodiment runs on a property clue mining platform. The property clue mining platform includes a text processing model and a clue classification model. The property clue mining device 300 includes:

[0155] The text extraction module 301 is used to receive a property clue mining instruction, obtain the case text of the pending property case, and obtain the pending case text;

[0156] Summary extraction module 302, used to import the case text to be processed into the property clue mining platform, extract summary information of the case text to be processed, and obtain summary text;

[0157] A semantic encoding module 303 is used to perform semantic encoding on the summary text using the word embedding matrix in the text processing model to obtain a semantic feature vector;

[0158] A local convolution module 304 is used to construct a text feature sequence based on the semantic feature vector, and use the convolution kernel in the clue classification model to perform a local convolution operation on the text feature sequence to obtain a local feature vector;

[0159] A vector concatenation module 305 is used to concatenate the semantic feature vector and the local feature vector to obtain a feature fusion vector;

[0160] The clue matching module 306 is used to determine the target property clue label that matches the property case to be processed based on the feature fusion vector, and output the target property clue label.

[0161] Furthermore, the property clue mining device 300 also includes a word embedding module, which is used to:

[0162] The first historical case text is obtained from a preset case database; keywords are extracted from the first historical case text, and the legal word segmentation text is screened from the keyword extraction results of the first historical case text; the legal word segmentation text is imported into the original word list of the preset text processing model; the original word list is vectorized to obtain a word embedding matrix.

[0163] Furthermore, the property clue mining device 300 further includes a first training module, which is used to:

[0164] The second historical case text is obtained from a preset case database, and data annotation is performed on the second historical case text; a training data set and a test data set are constructed based on the annotated second historical case text; a preset text processing model is pre-trained using the training data set; and a model performance test is performed on the pre-trained preset text processing model using the test data set.

[0165] Furthermore, the text processing model is a pre-trained language model based on the Transformer architecture. The Transformer architecture includes several encoding layer structures and several decoding layer structures. The first training module is specifically used for:

[0166] The historical case texts in the training data set are mapped into a vector representation through a word embedding matrix, and a training sample feature matrix is ​​constructed; the training sample feature matrix is ​​sequentially imported into several encoding layers of the Transformer architecture according to the feature hierarchy relationship to perform encoding operations on the training sample features; the feature encoding result output by the encoding layer is obtained, and the feature encoding result is input into the decoding layer to perform decoding operations on the encoded features; the feature decoding result is obtained, and the feature decoding result is compared with the first true label of the historical case text to obtain a first prediction error, wherein the first true label is the annotated text semantic label; the parameters of the Transformer architecture are iteratively updated based on the first prediction error until the model is fitted to obtain a preset text processing model.

[0167] Furthermore, the clue classification model is a pre-trained neural network model based on a text convolutional neural network architecture, and a plurality of convolution kernels are preset in the text convolutional neural network. The property clue mining device 300 also includes a second training module, which is used to:

[0168] The feature coding result output by the coding layer is parsed to obtain the feature coding vector output by the coding layer; a feature coding sequence is constructed based on the feature coding vector; an adapted target convolution kernel is determined among several convolution kernels in the preset clue classification model; a local convolution operation is performed on the feature coding sequence using the target convolution kernel to obtain a local coding vector; the feature coding vector and the local coding vector are concatenated to obtain a coding fusion vector; based on the feature fusion vector, a property clue label matching the historical case text is determined to obtain a historical property clue label; the historical property clue label is compared with a second true label of the historical case text to obtain a second prediction error, wherein the second true label is the annotated property clue label; the parameters of the clue classification model are iteratively updated based on the second prediction error until the model is fitted to obtain a preset text processing model.

[0169] Furthermore, the property clue mining device 300 further includes a vector screening module, which is used to:

[0170] The feature coding vectors are screened to obtain high-order feature coding vectors; and a feature coding sequence is constructed based on the high-order feature coding vectors.

[0171] Furthermore, the property clue mining device 300 also includes a third training module, and the third training module is used for:

[0172] Initial model weight parameters are configured for the text processing model and the clue classification model; a cross entropy loss function is set, and based on the first prediction error and the second prediction error, and with minimizing the cross entropy loss function as the optimization goal, the initial model weight parameters are dynamically updated to obtain the optimal model weight parameters; based on the optimal model weight parameters, the text processing model and the clue classification model are configured in the property clue mining platform.

[0173] The present application discloses a property clue mining device, which belongs to the field of artificial intelligence technology. By integrating the two models of text processing and clue classification in the property clue mining platform, the case text is first received, its summary information is extracted, and the word embedding matrix of the text processing model is used to convert the summary into a semantic feature vector, and then a text feature sequence is constructed based on the semantic feature vector, and the convolution kernel sequence of the clue classification model is applied to extract local features, and the local features are fused with the semantic features to form a comprehensive feature fusion vector, and finally the target property clue label related to the case is accurately matched and output based on the fusion vector. The present application effectively solves the dependence of traditional property clue mining on artificial feature engineering by automatically extracting case text summaries, generating semantic features using word embedding matrices, and capturing local features in combination with convolution operations, thereby reducing the difficulty and cost of operation. At the same time, the feature fusion strategy enhances the generalization ability of the model, reduces the dependence on data quality and sample size, and improves the accuracy and efficiency of property clue mining.

[0174] To solve the above technical problems, the present application also provides a computer device. Figure 4 , Figure 4 This is a basic structural block diagram of the computer device in this embodiment.

[0175] The computer device 4 includes a memory 41, a processor 42, and a network interface 43 that are interconnected through a system bus. It should be noted that the figure only shows a computer device 4 with components 41-43, 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), field-programmable gate arrays (Field-Programmab l eGate Array, FPGA), digital processors (Digital Signal Processor, DSP), embedded devices, etc.

[0176] 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.

[0177] The memory 41 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, disk, optical disk, etc. In some embodiments, the memory 41 can be an internal storage unit of the computer device 4, such as a hard disk or memory of the computer device 4. In other embodiments, the memory 41 can also be an external storage device of the computer device 4, such as a plug-in hard disk equipped on the computer device 4, a smart memory card (Smart Media Card, SMC), a secure digital

[0178] (Secure Digital, SD) card, flash card, etc. Of course, the memory 41 can also include both the internal storage unit of the computer device 4 and its external storage device. In this embodiment, the memory 41 is usually used to store the operating system and various application software installed on the computer device 4, such as computer-readable instructions of the property clue mining method. In addition, the memory 41 can also be used to temporarily store various data that have been output or are to be output.

[0179] The processor 42 may be a central processing unit (CPU), a controller, a microcontroller, a microprocessor, or other data processing chip in some embodiments. The processor 42 is generally used to control the overall operation of the computer device 4. In this embodiment, the processor 42 is used to run computer-readable instructions stored in the memory 41 or process data, such as computer-readable instructions for running the property clue mining method.

[0180] The network interface 43 may include a wireless network interface or a wired network interface. The network interface 43 is generally used to establish a communication connection between the computer device 4 and other electronic devices.

[0181] The present application discloses a computer device, which belongs to the field of artificial intelligence technology. By integrating the two models of text processing and clue classification in the property clue mining platform, the case text is first received, its summary information is extracted, and the word embedding matrix of the text processing model is used to convert the summary into a semantic feature vector, and then a text feature sequence is constructed based on the semantic feature vector, and the convolution kernel sequence of the clue classification model is applied to extract local features, and the local features are fused with the semantic features to form a comprehensive feature fusion vector, and finally the target property clue label related to the case is accurately matched and output based on the fusion vector. The present application automatically extracts the case text summary, generates semantic features using the word embedding matrix, and captures local features in combination with the convolution operation, which effectively solves the dependence of traditional property clue mining on artificial feature engineering and reduces the difficulty and cost of operation. At the same time, the feature fusion strategy enhances the generalization ability of the model, reduces the dependence on data quality and sample size, and improves the accuracy and efficiency of property clue mining.

[0182] The present application also provides another embodiment, 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 property clue mining method as described above.

[0183] The present application discloses a computer-readable storage medium, which belongs to the field of artificial intelligence technology. By integrating the two models of text processing and clue classification in the property clue mining platform, the case text is first received, its summary information is extracted, and the word embedding matrix of the text processing model is used to convert the summary into a semantic feature vector, and then a text feature sequence is constructed based on the semantic feature vector, and the convolution kernel sequence of the clue classification model is applied to extract local features, and the local features are fused with the semantic features to form a comprehensive feature fusion vector, and finally the target property clue label related to the case is accurately matched and output based on the fusion vector. The present application effectively solves the dependence of traditional property clue mining on artificial feature engineering by automatically extracting case text summaries, generating semantic features using word embedding matrices, and capturing local features in combination with convolution operations, thereby reducing the difficulty and cost of operation. At the same time, the feature fusion strategy enhances the generalization ability of the model, reduces the dependence on data quality and sample size, and improves the accuracy and efficiency of property clue mining.

[0184] 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.

[0185] 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.

[0186] 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 property clue mining method, characterized in that: The property clue mining method is applied to a property clue mining platform, the property clue mining platform includes a text processing model and a clue classification model, and the property clue mining method includes: Receive a property clue mining instruction, obtain the case text of the pending property case, and obtain the pending case text; Importing the pending case text into the property clue mining platform, extracting summary information of the pending case text, and obtaining a summary text; Using the word embedding matrix in the text processing model to semantically encode the summary text to obtain a semantic feature vector; Constructing a text feature sequence based on the semantic feature vector, and performing a local convolution operation on the text feature sequence using a convolution kernel in the clue classification model to obtain a local feature vector; Concatenating the semantic feature vector and the local feature vector to obtain a feature fusion vector; A target property clue label matching the property case to be processed is determined based on the feature fusion vector, and the target property clue label is output.

2. The property clue mining method according to claim 1, characterized in that: Before the step of using the word embedding matrix in the text processing model to semantically encode the summary text to obtain a semantic feature vector, the method further includes: Obtaining a first historical case text from a preset case database; Performing keyword extraction on the first historical case text, and screening legal segmentation text from the keyword extraction results of the first historical case text; Importing the legal segmentation text into an original vocabulary of a preset text processing model; The original word list is vectorized to obtain the word embedding matrix.

3. The property clue mining method according to claim 2, characterized in that: After the step of vectorizing the original word list to obtain the word embedding matrix, the method further includes: Acquire a second historical case text from the preset case database, and perform data annotation on the second historical case text; Constructing a training data set and a test data set based on the annotated second historical case text; Pre-training the preset text processing model using the training data set; The test data set is used to perform a model performance test on the pre-trained preset text processing model.

4. The property clue mining method according to claim 3, characterized in that: The text processing model is a pre-trained language model based on a Transformer architecture, wherein the Transformer architecture includes a plurality of encoding layer structures and a plurality of decoding layer structures, and the step of pre-training the preset text processing model using the training data set specifically includes: Mapping the historical case texts in the training data set into vector representations through the word embedding matrix, and constructing a training sample feature matrix; Importing the training sample feature matrix into several encoding layers of the Transformer architecture in sequence according to the feature hierarchy relationship to perform encoding operations on the training sample features; Obtaining the feature encoding result output by the encoding layer, and inputting the feature encoding result into the decoding layer to perform a decoding operation of the encoded feature; Obtaining a feature decoding result, and comparing the feature decoding result with a first true label of the historical case text to obtain a first prediction error, wherein the first true label is a marked text semantic label; The parameters of the Transformer architecture are iteratively updated based on the first prediction error until the model is fitted to obtain the preset text processing model.

5. The property clue mining method according to claim 4, characterized in that: The clue classification model is a pre-trained neural network model based on a text convolutional neural network architecture, and a plurality of convolution kernels are preset in the text convolutional neural network. Before the step of constructing a text feature sequence based on the semantic feature vector and performing a local convolution operation on the text feature sequence using the convolution kernel in the clue classification model to obtain a local feature vector, the method further includes: Parsing the feature coding result output by the coding layer to obtain the feature coding vector output by the coding layer; Constructing a feature coding sequence based on the feature coding vector; Determine an adapted target convolution kernel among several convolution kernels in a preset clue classification model; Using the target convolution kernel to perform a local convolution operation on the feature coding sequence to obtain a local coding vector; Concatenate the feature coding vector and the local coding vector to obtain a coding fusion vector; Determine a property clue label that matches the historical case text based on the feature fusion vector to obtain a historical property clue label; Comparing the historical property clue label with a second true label of the historical case text to obtain a second prediction error, wherein the second true label is the annotated property clue label; The parameters of the clue classification model are iteratively updated based on the second prediction error until the model is fitted to obtain the preset text processing model.

6. The property clue mining method according to claim 5, characterized in that: The step of constructing a feature coding sequence based on the feature coding vector specifically includes: Screening the feature coding vector to obtain a high-order feature coding vector; The feature coding sequence is constructed based on the high-order feature coding vector.

7. The property clue mining method according to claim 6, characterized in that: After the step of iteratively updating the parameters of the clue classification model based on the second prediction error until the model is fitted to obtain the preset text processing model, the method further includes: Configuring initial model weight parameters for the text processing model and the clue classification model; Setting a cross entropy loss function, and based on the first prediction error and the second prediction error, dynamically updating the initial model weight parameters with minimizing the cross entropy loss function as an optimization goal, to obtain optimal model weight parameters; The text processing model and the clue classification model are configured in the property clue mining platform based on the optimal model weight parameters.

8. A property clue mining device, characterized in that: The property clue mining device runs on a property clue mining platform, the property clue mining platform includes a text processing model and a clue classification model, and the property clue mining device includes: A text extraction module, used to receive a property clue mining instruction, obtain the case text of the pending property case, and obtain the pending case text; A summary extraction module, used for importing the case text to be processed into the property clue mining platform, extracting summary information of the case text to be processed, and obtaining a summary text; A semantic encoding module, used to perform semantic encoding on the summary text using the word embedding matrix in the text processing model to obtain a semantic feature vector; A local convolution module, used for constructing a text feature sequence based on the semantic feature vector, and performing a local convolution operation on the text feature sequence using the convolution kernel in the clue classification model to obtain a local feature vector; A vector concatenation module, used for concatenating the semantic feature vector and the local feature vector to obtain a feature fusion vector; A clue matching module is used to determine a target property clue label that matches the property case to be processed based on the feature fusion vector, and output the target property clue label.

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 property clue mining method according to 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 a processor, the steps of the property clue mining method according to any one of claims 1 to 7 are implemented.

Images