Information identification method, device, equipment, medium and product

Abstract

The present disclosure provides an information recognition method, device, equipment, medium and product, relates to the technical field of artificial intelligence, and can be applied to the technical field of financial technology. The method comprises the following steps: acquiring text data to be recognized; pre-processing the text data to be recognized to obtain first text features. The first text features are input into a trained recognition model to recognize target information included in the text data to be recognized. The recognition model is obtained by training a spatial transformation network model based on domain adaptation and using a heterogeneous transfer learning manner. Recognition labels of the recognition model are configured according to business attribute information. The information recognition method, device, equipment, medium and product can accurately recognize target information in the text data to be recognized.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the technical field of artificial intelligence, can be applied to the technical field of financial technology, and in particular to an information recognition method and device, equipment, medium and product. BACKGROUND

[0002] With the in-depth development of deep learning technology, natural language processing has also entered a stage of vigorous development. Nowadays, various natural language processing technologies have been used in production and work to greatly reduce labor costs. With the rapid development of financial service business, various service businesses have emerged, such as customer voice, a natural language processing technology. Customer voice recognizes the voice of customer calls to enable banks to provide services to customers according to the recognition results.

[0003] However, the existing customer voice has low recognition accuracy for the text data of customer calls, which affects the accurate determination of the demands of customers by institutions, thereby affecting the accuracy of services provided by institutions to customers and reducing customer experience. SUMMARY

[0004] Therefore, the main purpose of the present disclosure is to provide an information recognition method, device, equipment, medium and product, which aims to at least partially solve the technical problem of low recognition accuracy of the existing customer voice for the text data of customer calls.

[0005] To achieve the above-mentioned purpose, a first aspect of an embodiment of the present disclosure provides an information recognition method, comprising: obtaining text data to be recognized; preprocessing the text data to be recognized to obtain first text features; inputting the first text features into a trained recognition model to recognize target information included in the text data to be recognized, wherein the recognition model is obtained by training a spatial transformation network model in a heterogeneous transfer learning manner based on domain adaptation, and the recognition label of the recognition model is configured according to business attribute information.

[0006] According to an embodiment of the present disclosure, the information recognition method further comprises: sequentially performing dimension reduction and dimension increase on the first text features to filter invalid information in the first text features to obtain second text features; inputting the first text features into the trained recognition model to recognize the target information included in the text data to be recognized.

[0007] According to an embodiment of the present disclosure, the preprocessing of the text data to be recognized to obtain the first text features comprises: determining the length of each text in the text data; comparing the length of each text with a preset length threshold to determine texts with a length greater than the preset length threshold; and cutting a part from the texts with a length greater than the preset length threshold to obtain the first text features of a preset dimension.

[0008] According to an embodiment of the present disclosure, the information identification method further comprises: training the identification model, comprising: obtaining historical text data; determining a training data set and a test data set based on the historical text data; labeling the training data set according to the identification label, and labeling part of the test data set according to the identification label; inputting the labeled training data set and the labeled part of the test data into the space transformation network model for heterogeneous transfer learning with a statistical distance between a source domain where the training data set is located and a target domain where the test data set is located as a loss function, to obtain a trained identification model.

[0009] According to an embodiment of the present disclosure, the determining a training data set and a test data set based on the historical text data comprises: down-sampling part of the historical text data to obtain the training data set; and randomly sampling another part of the historical text data to obtain the test data set.

[0010] According to an embodiment of the present disclosure, before inputting the labeled training data set and the labeled part of the test data into the space transformation network model for heterogeneous transfer learning, the method further comprises: mapping the training data set and the test data set to the same dimension.

[0011] According to an embodiment of the present disclosure, the training data set and the test data set are mapped to a reproducing kernel Hilbert space so that the training data set and the test data set are in the same dimension.

[0012] According to an embodiment of the present disclosure, the identification label comprises: an account query label, a bank and outlet information query label, an account and debit card label, a transfer and remittance label, a personal credit label, a deposit label, a financial product label, a fund label, a precious metal label, a personal mobile bank label, a personal online bank label, a self-service machine label, a messenger label, and a comprehensive label, wherein the comprehensive label is used to represent a label corresponding to business attribute information other than account query, bank and outlet information query, account and debit card, transfer and remittance, personal credit, deposit, financial product, fund, precious metal, personal mobile bank, personal online bank, self-service machine, and messenger.

[0013] A second aspect of the embodiments of the present disclosure provides an information recognition device, comprising: an acquisition module configured to acquire text data to be recognized; a preprocessing module configured to preprocess the text data to be recognized to obtain first text features; and a recognition module configured to input the first text features into a trained recognition model to recognize target information included in the text data to be recognized, wherein the recognition model is obtained by training a spatial transformation network model based on domain adaptation and using heterogeneous transfer learning, and recognition labels of the recognition model are configured according to business attribute information.

[0014] A third aspect of the embodiments of the present disclosure provides an electronic device, comprising: one or more processors; and a storage device configured to store one or more programs, wherein the one or more programs, when executed by the one or more processors, cause the one or more processors to perform the information recognition method described above.

[0015] A fourth aspect of the embodiments of the present disclosure provides a computer-readable storage medium having stored thereon executable instructions that, when executed by a processor, cause the processor to perform the information recognition method described above.

[0016] A fifth aspect of the embodiments of the present disclosure provides a computer program product comprising a computer program, which, when executed by a processor, implements the information recognition method described above.

[0017] The information recognition method, device, equipment, medium and product provided by the embodiments of the present disclosure have at least the following beneficial effects:

[0018] Since the recognition labels are configured according to the attribute information of the business, and the recognition model is obtained by training the spatial transformation network model based on domain adaptation and using heterogeneous transfer learning, the recognition model can better adapt to information recognition in specific business scenarios, thereby improving the accuracy of target information recognition.

[0019] Since the text features are first processed by dimensionality increasing and then by dimensionality reduction, the invalid information in the first text features can be filtered out, on the one hand, the influence of invalid data on target information recognition in the recognition process is reduced, and the accuracy of target information recognition is improved, on the other hand, the amount of data involved in the recognition process is reduced, and the recognition efficiency is improved.

[0020] Since the text length threshold is set to truncate the text with too long length in the first text features, the amount of data for calculation is reduced on the premise of ensuring that as much valid data as possible is truncated, and the recognition efficiency is further improved.

[0021] In the training process of the identification model, since all the training data sets and part of the test data sets are labeled based on the identification label, and part of the labeled data of the target domain is used as the training set at the same time as the source domain, and the statistical distance between the source domain where the training data set is located and the target domain where the test data set is located is used as the loss function for training, therefore, the difference between the source domain and the target domain can be better reduced, the domain adaptation can be realized, and therefore the problem of model failure caused by the fact that not all the source domain and the target domain have the same characteristics in the training process can be solved, thereby improving the accuracy of target information identification.

[0022] In the process of obtaining the training data set and the test data set, since the data in the historical text data is down-sampled, the obtained training data of each category is balanced, therefore, the problem of low accuracy of target information identification caused by the fact that the data distribution of the accuracy and the result of the model training tends to some categories can be avoided, thereby improving the accuracy of target information identification. Since the data in the historical text data is randomly sampled, the test data obtained can cover the text data involved in the current business as much as possible, therefore, the identification model trained is more comprehensive, thereby making the application scope of the identification model more comprehensive, and further improving the accuracy of the target information.

[0023] Since the training data set and the test data set are mapped to the same dimension, the problem of model failure caused by the fact that not all the source domain and the target domain have the same characteristics in the training process can be solved, thereby further improving the accuracy of target information identification.

[0024] By defining the specific type of the identification label, the identification method can be better applied to the field of financial service technology. BRIEF DESCRIPTION OF DRAWINGS

[0025] In order to more clearly illustrate the technical solutions in the embodiments of the present disclosure or the prior art, the drawings needed to be used in the embodiments or the prior art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present disclosure, and other drawings can be obtained by those skilled in the art without creative labor based on the drawings shown.

[0026] Figure 1 The system architecture 100 of the information identification method and device according to an embodiment of the present disclosure is schematically shown;

[0027] Figure 2 The flowchart of the information identification method according to an embodiment of the present disclosure is schematically shown;

[0028] Figure 3 The flowchart of the information identification method according to another embodiment of the present disclosure is schematically shown;

[0029] Figure 4 schematically illustrates a flowchart of training the recognition model according to an embodiment of the present disclosure; Figure 1 schematically illustrates a flowchart of pre-processing the text data to be recognized in operation S202 shown;

[0030] Figure 5 schematically illustrates a flowchart of training the recognition model according to an embodiment of the present disclosure;

[0031] Figure 6 schematically illustrates a flowchart of training the recognition model according to an embodiment of the present disclosure; Figure 5 schematically illustrates a flowchart of determining the training data set and the test data set based on the historical text data in operation S502 shown;

[0032] Figure 7A schematically illustrates a curve graph of the AUC corresponding to the training data set and the test data set respectively according to an embodiment of the present disclosure;

[0033] Figure 7B schematically illustrates a curve graph of the loss function according to an embodiment of the present disclosure;

[0034] Figure 8 schematically illustrates a confusion matrix graph of the result on the labeled test set in the STN model for the validation set with epoch 40 according to an embodiment of the present disclosure;

[0035] Figure 9 schematically illustrates a labeling result graph of labeling the data of the test set according to an embodiment of the present disclosure

[0036] Figure 10 schematically illustrates a block diagram of an information recognition apparatus according to an embodiment of the present disclosure;

[0037] Figure 11 schematically illustrates a block diagram of an information recognition apparatus according to another embodiment of the present disclosure;

[0038] Figure 12 schematically illustrates a block diagram of an information recognition apparatus according to yet another embodiment of the present disclosure;

[0039] Figure 13 schematically illustrates a block diagram of an electronic device suitable for implementing the method described above according to an embodiment of the present disclosure. DETAILED DESCRIPTION

[0040] Hereinafter, embodiments of the present disclosure will be described with reference to the accompanying drawings. It is to be understood, however, that the description is merely exemplary of the present disclosure, and is not intended to limit the scope of the present disclosure. In the following detailed description of the embodiments of the present disclosure, numerous specific details are set forth in order to provide a thorough understanding of the present disclosure. However, it will be apparent to one skilled in the art that the present disclosure can be practiced without these specific details. In other instances, well-known structures and functions have not been described in detail in order to avoid obscuring aspects of the present disclosure.

[0041] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the present disclosure. As used herein, the term "including" "comprising" and the like are meant to be inclusive, and are intended to mean that there are additions to the claimed elements or steps, not exclusion, with or without a limitation.

[0042] All terms used herein including technical and scientific terms have the meanings as commonly understood by one of ordinary skill in the art unless otherwise defined. It should be noted that the terms used herein are defined as having meanings consistent with the context of the present description, and should not be interpreted in an idealized or overly formal way.

[0043] In situations where similar terminology is used in a similar manner, one of ordinary skill in the art will be aware that the terminology is to be interpreted in the manner in which it is commonly understood by those skilled in the art (e.g., "a system having at least one of A, B, and C" should be interpreted to mean a system having A alone, a system having B alone, a system having C alone, a system having both A and B, a system having both A and C, a system having both B and C, and / or a system having A, B, and C, etc.). In situations where similar terminology is used in a similar manner, one of ordinary skill in the art will be aware that the terminology is to be interpreted in the manner in which it is commonly understood by those skilled in the art (e.g., "a system having at least one of A, B, or C" should be interpreted to mean a system having A alone, a system having B alone, a system having C alone, a system having both A and B, a system having both A and C, a system having both B and C, and / or a system having A, B, and C, etc.).

[0044] Some of the blocks and / or combinations of the blocks in the flowcharts can be implemented by computer program instructions. Such computer program instructions can be provided to a processor of a general purpose computer, special purpose computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions / acts specified in the flowcharts block or blocks. The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operations to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus implement the functions / acts specified in the flowcharts block or blocks.

[0045] In the technical solutions of the present disclosure, the collection, storage, use, processing, transmission, provision, disclosure and application of information comply with relevant laws and regulations, necessary security measures are taken, and the public order and good customs are not violated.

[0046] In the technical solutions of the present disclosure, if user personal information needs to be obtained, the user's authorization or consent is obtained before the user personal information is obtained or collected.

[0047] To solve the technical problems in the related art, the present embodiment provides an information recognition method, which includes: obtaining text data to be recognized; preprocessing the text data to be recognized to obtain first text features. The first text features are input into a trained recognition model to recognize target information included in the text data to be recognized, wherein the recognition model is obtained by training a spatial transformation network model based on domain adaptation and using heterogeneous transfer learning, and the recognition labels of the recognition model are configured according to business attribute information.

[0048] Figure 1 The system architecture 100 of the information recognition method and device according to the present embodiment is schematically shown. It should be noted that, Figure 1 The shown is only an example of the system architecture to which the present embodiment can be applied, to help those skilled in the art understand the technical content of the present disclosure, but does not mean that the present embodiment cannot be used in other devices, systems, environments or scenarios.

[0049] As Figure 1 shown, the system architecture 100 according to the embodiment can include a communication device 101, a storage device 102, a network 103 and a server 104. The network 103 is used to provide a communication link between the communication device 101, the storage device 102 and the server 104.

[0050] The communication device 101 may, for example, be an electronic device with a display screen or a touch display screen, and may, for example, include a telephone, a mobile phone, a computer, etc. The communication device 101 can be used to receive incoming call voice data of a user and convert the voice data into text data.

[0051] The storage device 102 can be a hardware storage device or a software storage device, and the present disclosure does not limit it. The storage device 102 is used to store the incoming call voice data of the user received by the communication device 101 and the text data converted from the incoming call voice data.

[0052] The network 103 can include various connection types, such as wired, wireless communication links, or fiber optic cables, and the like. The wired mode can be, for example, connected by using a cable and any one of the following various interfaces: fiber channel, infrared interface, D-type data interface, serial interface, USB interface, USB Type-C interface, or Dock interface. The wireless mode can be, for example, connected by using wireless communication, which can use any one of a plurality of wireless technology standards such as Bluetooth, Wi-Fi, Infrared, ZigBee, and the like.

[0053] The server 104 receives the text data to be recognized currently acquired by the communication device 101 through the network 103, pre-processes the text data to be recognized to obtain first text features, and then inputs the first text features into a trained recognition model stored on the server 104 to recognize target information included in the text data to be recognized. In the process of training the recognition model, the server 104 acquires historical text data from the storage device 102 through the network 103, determines a training data set and a test data set based on the historical text data, labels the training data set according to a recognition label, labels part of the test data of the test data set according to the recognition label, and takes the statistical distance between the source domain where the training data set is located and the target domain where the test data set is located as a loss function. The labeled training data set and the labeled part of the test data are input into a spatial transformation network model for heterogeneous transfer learning to obtain the trained recognition model.

[0054] It should be noted that the information recognition method provided by the embodiments of the present disclosure can be executed by the server 104. Accordingly, the information recognition apparatus provided by the embodiments of the present disclosure can be arranged in the server 104. Alternatively, the information recognition method provided by the embodiments of the present disclosure can also be executed by a server or a server cluster different from the server 104 and capable of communicating with the communication device 101 and / or the storage device 102 and / or the server 104. Accordingly, the information recognition apparatus provided by the embodiments of the present disclosure can also be arranged in a server or a server cluster different from the server 104 and capable of communicating with the communication device 101 and / or the storage device 102 and / or the server 104. Alternatively, the information recognition method provided by the embodiments of the present disclosure can also be partially executed by the server 104, partially executed by the communication device 101, and partially executed by the storage device 102. Accordingly, the information recognition apparatus provided by the embodiments of the present disclosure can also be partially arranged in the server 104, partially arranged in the communication device 101, and partially arranged in the storage device 102.

[0055] It should be understood that Figure 1 The number of communication devices, storage devices, networks, and servers in the above description is only illustrative. According to the needs of implementation, there can be any number of communication devices, storage devices, networks, and servers.

[0056] The information recognition method provided by the embodiments of the present disclosure can be applied to the field of financial technology. For example, for a bank, it is particularly important to provide precise services for customers in order to further tap the value of customers at different levels and attract more high-quality customers. For example, banks generally have a special customer complaint department to provide precise services for customer calls, but due to the large base of customer groups and the limited number of staff in the customer complaint department, in order to reduce labor costs and improve customer service efficiency, the technology of using chat robots to process customer call voice data or text data has gradually been applied. In this scenario, how to accurately identify the target information in the call voice data or text data is the key to providing point-by-point precise services for customers. If the target information in the call voice data or text data can be accurately identified, it can be quickly allocated to the relevant departments involved in the subsequent services, such as retail / operating departments, and even refined to customer managers, so that the bank can provide one-on-one services for the calling customers. The information recognition method provided by the embodiments of the present disclosure can achieve accurate identification of the target information in the call voice data or text data.

[0057] It should be understood that the information recognition method provided by the embodiments of the present disclosure is not limited to application in the field of financial technology, but can also be used in any field other than the financial field. The above description is only exemplary, and the information recognition method of the embodiments of the present disclosure can be applied to any field involving identification of target information in call voice data or text data, such as e-commerce, logistics and other technical fields.

[0058] The following will be based on Figure 1 a description of the scenario of information recognition, through Figures 2-9 a detailed description of the information recognition method of the embodiments of the present disclosure.

[0059] Figure 2 An illustrative flowchart of an information recognition method according to an embodiment of the present disclosure is shown.

[0060] As Figure 2 shown, the information recognition method may, for example, include operations S201-S203.

[0061] In operation S201, text data to be recognized is obtained.

[0062] In the embodiments of the present disclosure, the user may currently feedback voice data containing a complaint through a telephone, or manually input text data containing a complaint through a mobile phone bank or an online bank. In order to facilitate the use of natural language processing data to identify target information, the server can convert the voice data into text data.

[0063] In operation S202, the text data to be identified is preprocessed to obtain first text features.

[0064] In an embodiment of the present disclosure, the text data to be identified contains different lengths of text. Considering that due to the limitation of computing power, a large-scale recognition model cannot be used to recognize the text data, the length of the text data to be identified can be processed in a pre-processing manner.

[0065] In operation S203, the first text features are input into the trained recognition model to identify the target information included in the text data to be identified.

[0066] In an embodiment of the present disclosure, the recognition model is trained based on domain adaptation and heterogeneous transfer learning of a spatial transformation network model, and the recognition label of the recognition model is configured according to the attribute information of the business.

[0067] It can be understood that different text data contains different target information. For example, the text data is “I want to query the information of the bank of deposit”, and the target information contained can be “query the bank of deposit”. The text data is “I want to borrow money”, and the target information contained can be “personal credit”. Therefore, in order to accurately identify various types of text data, the recognition model needs to have high adaptability.

[0068] According to an embodiment of the present disclosure, since the recognition label is configured according to the attribute information of the business, and the recognition model is trained based on domain adaptation and heterogeneous transfer learning of a spatial transformation network model, the text data to be identified is recognized by the recognition model, so that the recognition model is more suitable for information recognition in a specific business scenario, thereby improving the accuracy of target information recognition.

[0069] Figure 3 A flowchart of an information recognition method according to another embodiment of the present disclosure is schematically shown.

[0070] As shown in Figure 3 , the information recognition method may, for example, further include operations S301-S303.

[0071] In operation S301, the first text features are sequentially reduced in dimension and increased in dimension to filter out invalid information in the first text features, to obtain second text features.

[0072] In operation S302, the first text features are input into the trained recognition model to identify the target information included in the text data to be identified.

[0073] For example, the server currently receives voice data sent by a user, which contains some environmental noise (e.g., the sound of vehicles driving on the road, the sound of the wind blowing, the sound of rain, etc.). The existence of this part of environmental noise increases the computational load of subsequent information recognition, reduces the recognition efficiency, and may also affect the accuracy of target information recognition. For another example, the voice data or text data expressed by the user to express his / her appeal is not concise, and there are repeated sentences or some information that is not closely related to the final appeal. This information does not positively affect the identification of target information, but increases the computational load and reduces the recognition efficiency.

[0074] According to an embodiment of the present disclosure, since the processing mode of first increasing the dimension and then reducing the dimension of the text feature is adopted, the invalid information in the first text feature can be filtered out. On the one hand, the influence of invalid data on target information recognition in the recognition process is reduced, and the accuracy of target information recognition is improved. On the other hand, the amount of data involved in the recognition process is reduced, and the recognition efficiency is improved.

[0075] Figure 4 An example of the operation S202 of preprocessing the text data to be identified according to an embodiment of the present disclosure is shown. Figure 1 The flowchart of the operation S202 of preprocessing the text data to be identified is shown.

[0076] As shown in the flowchart of the operation S202 of preprocessing the text data to be identified, the operation S202 can include operations S401-S403. Figure 4

[0077] In operation S401, the length of each text in the text data is determined.

[0078] In operation S402, the length of each text is compared with a preset length threshold value, and the text whose length is greater than the preset length threshold value is determined.

[0079] In operation S403, a part of the text whose length is greater than the preset length threshold value is cut off to obtain a first text feature of a preset dimension.

[0080] In an embodiment of the present disclosure, the length of the text data can be understood as the number of characters contained in the text data. The preset length threshold value can be determined according to the computing capacity of the recognition model, for example, the preset length threshold value is 10 characters, 15 characters, etc. It should be immediately understood that the value of the aforementioned preset length threshold value is only exemplary and does not limit the present disclosure.

[0081] According to an embodiment of the present disclosure, since the text whose length is too long in the first text feature is cut off by setting the text length threshold value, the amount of data for calculation is reduced under the premise of ensuring that as much valid data as possible is cut off, and the recognition efficiency is further improved.

[0082] ​Further, on the basis of the above-mentioned embodiments, the information recognition method can further include training the recognition model.

[0083] Figure 5 A flowchart of training the recognition model according to an embodiment of the present disclosure is schematically shown.

[0084] As Figure 5 shown, the training of the recognition model can include operations S501-S504, for example.

[0085] In operation S501, historical text data is acquired.

[0086] In operation S502, a training data set and a test data set are determined based on the historical text data.

[0087] In operation S503, the training data set is labeled according to a recognition label, and part of the test data of the test data set is labeled according to the recognition label.

[0088] In operation S504, the labeled training data set and the labeled part of the test data are input into a spatial transformation network model to perform heterogeneous transfer learning with a statistical distance between a source domain where the training data set is located and a target domain where the test data set is located as a loss function, to obtain a trained recognition model.

[0089] In an embodiment of the present disclosure, acquiring historical text data can include acquiring historical work order data, and extracting historical text data from the historical work order data.

[0090] In an embodiment of the present disclosure, the spatial transformation network model (STN) can be constructed based on artificial intelligence technologies such as BERT (Bidirectional Encoder Representations from Transformers) model transfer learning and domain adaptation, for example. According to the potential value type of the incoming customer, the business types are re-divided into 14 categories, including account inquiry, branch information inquiry, account and debit card, transfer remittance, personal credit, deposit, finance, fund, precious metal, personal mobile bank, personal online bank, self-service machine, BOC messenger, and other categories. The automatic recognition process of each work order according to the business type is realized, and business support is provided for further mining of value customers of different levels of the bank. Details will be described later.

[0091] BERT is a deep bidirectional Transformer pre-training model. As an unsupervised learning, BERT can be trained using only a corpus of plain text. Similar to Transformer, BERT is also a deep network model that connects context, but it only uses the Encoder part, which is built by multiple layers of Encoders. Transformer is a deep learning neural network model for translation model, which retains the Encoder and Decoder modules. The Encoder is divided into two modules, Self-Attention and a layer of feedforward neural network layer. The Decoder adds an Encoder-Decoder Attention layer after the Self-Attention layer based on the Encoder.

[0092] BERT has made changes in input compared to Transformer. In Transformer, position information can be trained or directly used as a function, while in BERT, position information is directly obtained by training. Segment vector is like judging whether two texts are similar. This embedding has only two elements, 0 and 1. It will set the same word vector part of two similar sentences to 0, and the different part to 1. Finally, add the three to get the input of BERT.

[0093] In the training process, BERT will undergo two stages of training. The first stage is pre-training (Pre-Training), which is divided into two steps. The first step is to cover a certain proportion of words, so that the model can predict the covered words according to the context. The second step is to train continuous parameters by selecting several pairs of sentences. Some of these sentences are continuous, and some are not. Therefore, the model at this time is a binary classification model that determines which sentences are continuous and which are not. The second part of the stage, Fine-Tuning, can adapt different neural networks according to different tasks, such as classification model, prediction model, and even clustering and other unsupervised learning.

[0094] From the architecture of the model, the advantage of BERT over Transformer is that BERT overcomes the limitation of one-way, that is, BERT fuses the context on the left and right sides, thereby actually pre-training a bidirectional Transformer Encoder model. In actual use, BERT has shown better performance than other models, such as Transformer and Bi-LSTM, on many data sets. In the model of the embodiments of the present disclosure, the pre-trained model is used to retrain the parameters to adapt to the text, and the final architecture is a multi-layer full connection layer and a Decoder layer.

[0095] Further, in the process of implementing the present disclosure, the applicant finds that in supervised learning, the model obtained by the training set will not perform well on the test set due to the difference in data distribution between the training set and the test set. For example, in a graph classification model, if the content of the training set is daytime content, the RGB value of the image will be higher due to the problem of light, and if the test set is at night, the RGB value will be lower, thereby forming a situation in which light causes many algorithms to fail in the result. The same situation also occurs in recurrent neural network tasks, such as translation models, which have a large generalization error due to differences in language types and different usage methods. The training of the recognition model provided in the embodiments of the present disclosure also has corresponding problems.

[0096] Therefore, the embodiments of the present disclosure introduce domain adaptation (DA) in the training of the recognition model. Domain adaptation can be understood as a method for reducing generalization error. The core idea is to reduce the domain distance (discrepancy) between the domain (source domain) of the training set and the domain (target domain) of the test set through transformation, so that the trained model better adapts to the data of the test set.

[0097] The method of realizing domain adaptation is usually divided into two categories: reducing the domain distance as an optimization target in the training process and using adversarial learning to generate and confront data to achieve the target. Among them, domain distance reduction is a commonly used and intuitive method. The domain distance is the statistical distance between the domain of the test set and the domain of the training set. In domain adaptation, the maximum mean discrepancy (MMD) is used to measure it, which can be defined as formula (1):

[0098]

[0099] wherein, is a continuous function on the sample space, is the sample data in the test set in the continuous function the mean of the function values of the continuous function for the sample data in the training set the mean of the function values of the continuous function

[0100] Due to the arbitrariness of the continuous function It is difficult to directly calculate it, for example, therefore, before performing the operation of inputting the labeled training data set and the labeled part of the test data into the space transformation network model for heterogeneous transfer learning in operation S504, the method further comprises: mapping the training data and the test data to the same dimension.

[0101] Preferably, the training data set and the test data set can be mapped to a reproducing kernel Hilbert space (RKHS) so that the training data set and the test data set are in the same dimension. Specifically, the RKHS has the characteristic that any function in the space and its dependent variable can be expressed as the inner product of a set of kernel functions. In this way, the calculation method of MMD can be formula (2):

[0102]

[0103] wherein, is a Gaussian kernel (GBF kernel), i.e. formula (3):

[0104]

[0105] It should be noted that there are two reasons for using the Gaussian kernel here, one is that it can compress the norm of the function set () to (0, 1] through the exponential function in the positive interval (0, 1] to create conditions for the last step of formula (2), and the other is that it is equivalent to the Fourier transform on a non-countable set, which maps the data to an orthogonal space, making the domain have obvious advantages in classification problems.

[0106] Therefore, when actually calculating MMD, it can be calculated by Considering that formula (2) maps the data to RKHS in use, in order to achieve such mapping at the same time and ensure the correctness of the kernel function in use, the square of the Gaussian kernel is used in use, i.e. formula (4):

[0107]

[0108] wherein, the matrix K is obtained by the kernel function k(·,·), and L is a piecewise constant, and tr is the trace of the matrix, i.e. formula (5):

[0109]

[0110] Thus, the difference between the two domains can be calculated by formula (4), and the statistical distance between the two can be reduced by shortening the obtained MMD.

[0111] For the use of MMD, it is usually placed after the bottleneck layer in the fully connected layer, and the domain distance of the extracted features is used to achieve the optimal solution, which is obtained by fine tuning. Numerically, the value calculated by MMD is usually introduced as part of the loss function, and the value of the original loss function is linearly combined after introducing the hyperparameter, so as to reduce the domain distance and improve the generalization

[0112] In the embodiment of the present disclosure, for the current customer voice dataset, it is not completely the same as the test set, which highlights the necessity of transfer learning. In addition, different dialects will involve different word vectors, for example, the dialect "what happened", which may contain 3 or 4 words, and due to different pronunciation, it will result in different characters when converted to text, so it can be converted by hyterogeneous domain adaptation to solve the problem that not all source domains and target domains have the same features, thereby improving the generalization ability. Therefore, the STN model with high accuracy in transfer learning is used in the present disclosure. In transfer learning, the existing knowledge is called source domain, and the new knowledge to be learned is called target domain. Transfer learning is to migrate the knowledge of the source domain to the target domain. In the embodiment of the present disclosure, the source domain is the domain where the training set dataset is located, which is different from the test dataset, but has rich supervised information. The target domain is the domain where the test set dataset is located, which has a small amount of labels.

[0113] In the embodiment of the present disclosure, the loss function of the STN model is composed of two parts, Margin MMD (M m ) and Conditional MMD (M c ), and their expressions are formula (6) and formula (7) respectively:

[0114]

[0115] Among them, MMD is defined above, and in , C is the number of categories, and are the number of source domain data and the number of category k respectively, With respectively the number of target domain labeled data and the number of the number under the category k, is the number of target domain unlabeled data. And is defined as formula (8):

[0116]

[0117] Where R is the total number of cycles, r is the current cycle number, and y is the pseudo label generated by the model using its own parameters. The advantage of this recognition model is to abandon the original MMD's defect of ignoring the category, and by adding , the recognition model can gradually increase the weight of these added pseudo labels, thereby approximately achieving the effect of semi-supervised learning.

[0118] Therefore, based on the recognition model, the data of the customer's voice can be used as a training set to manually annotate part of the data, and then train the original unlabeled model to improve the accuracy.

[0119] Further, after the recognition model obtains the feature vector of the text data, it is necessary to analyze the feature vector. Similar to multiple fully connected layers in the field of computer vision (CV), the features are filtered through such fully connected layers. For example, different accents, dialects and regional dialogue features will have different effects on information, so the present disclosure adds STN domain adaptation in this part and gives its loss function formula (9):

[0120]

[0121] The best position for domain adaptation is to add a bottleneck, which means that the fully connected layer of the embodiment of the present disclosure needs to first reduce the dimension and then increase the dimension to filter a part of the invalid or small-impact feature information. After such a bottleneck layer, the domain adaptation layer is added, and the number of layers and the dimension are determined through fine-tuning. After completing the domain adaptation, the dimension is reduced to the probability value established before the embodiment of the present disclosure. Finally, the type is confirmed by the probability value, and the loss function involved is defined as formula (10):

[0122]

[0123] Where, Z is the final label obtained in training, Z true is the label according to the true annotation.

[0124] Therefore, the final loss function of the recognition model is formula (11):

[0125]

[0126] wherein, is a hyperparameter.

[0127] According to the embodiments of the present disclosure, in the training process of the identification model, since the entire training data set and part of the test data set are labeled based on the identification label, part of the labeled data of the target domain is used as the training set at the same time as the source domain, and the statistical distance between the source domain where the training data set is located and the target domain where the test data set is located is used as the loss function for training, therefore, the difference between the source domain and the target domain can be better reduced, the domain adaptation is realized, and therefore the problem of model failure caused by the fact that not all source domains and target domains have the same characteristics in the training process can be solved, thereby improving the accuracy of target information identification. Since the training data set and the test data set are mapped to the same dimension, the problem of model failure caused by the fact that not all source domains and target domains have the same characteristics in the training process can be solved, thereby further improving the accuracy of target information identification.

[0128] Figure 6 An example of a process of determining a training data set and a test data set based on historical text data is shown in operation S502 according to an embodiment of the present disclosure. Figure 5 An example of a process of determining a training data set and a test data set based on historical text data is shown in operation S502 according to an embodiment of the present disclosure.

[0129] As shown in Figure 6 The operation S502 of determining a training data set and a test data set based on historical text data may, for example, include operations S601-S602.

[0130] In operation S601, a part of the data in the historical text data is down-sampled to obtain the training data set.

[0131] In operation S602, another part of the data in the historical text data is randomly sampled to obtain the test data set.

[0132] In an embodiment of the present disclosure, since the customer call voice text data provided does not provide classification labels, and due to the limitations of hardware and software facilities, it is difficult to directly use the data set for classification or clustering operation.

[0133] Based on this, an embodiment of the present disclosure selects the full amount of customer voice data as the training set of the initial pre-training model, selects work_order_biz_class_cd, work_order_biz_sub_class_cd, and cust_opinion_idtfy_cd as training labels, and selects work_order_cust_problem_desc as the training data set for training.

[0134] However, since the content of the original label does not correspond, embodiments of the present disclosure reconfigure the training label (referred to as the identification label when identifying) according to the attribute information of the business (for example, can be according to the customer voice work order category and the mapping relationship between the selected topic label category is given in Table 6.1:

[0135] Table 6.1

[0136]

[0137] As shown in Table 6, the identification label includes: an account query label, an account opening bank and branch information query label, an account and debit card label, a transfer remittance label, a personal credit label, a deposit label, a financial label, a fund label, a precious metal label, a personal mobile bank label, a personal online bank label, a self-service machine label, a messenger label, and a comprehensive label. The comprehensive label is used to represent the label corresponding to the business attribute information other than account query, account opening bank and branch information query, account and debit card, transfer remittance, personal credit, deposit, financial, fund, precious metal, personal mobile bank, personal online bank, self-service machine, and messenger. By defining the specific type of the identification label, the identification method can be better applied to the field of financial services.

[0138] It should be noted that the amount of data on each category is not the same, for example, the proportion of account and debit card in customer voice is very large, which will lead to unbalanced data set samples, so that the correct rate of training and the data distribution of the result tend to this category.

[0139] For example, the customer voice work order data set is very large, more than 13 million, and complete training of data on time is not allowed, so the customer voice work order data is down-sampled to balance the amount of data in each category. Embodiments of the present disclosure, for example, select 140,000 data, that is, 10,000 data per category for training. After obtaining the initial training set, the provided customer call text data is also randomly sampled, and multiple rows of data are extracted for manual label annotation. A small data set is used for testing set, and multiple rows of data can also be extracted for manual label annotation as a validation set. For example, embodiments of the present disclosure randomly sample 269 million customer call text data provided, extract 510 rows of data for manual label annotation, and use a small data set for the validation set.

[0140] According to the embodiments of the present disclosure, in the process of obtaining the training data set and the test data set, the training data of each category obtained is balanced due to the undersampling of the data in the historical text data, so that the problem that the accuracy of target information recognition is low due to the data distribution of the accuracy and results of model training tending to some categories can be avoided, thereby improving the accuracy of target information recognition. Since the data in the historical text data is randomly sampled, the test data obtained can cover the text data involved in the current business as much as possible, so that the recognition model trained is more comprehensive, thereby making the application scope of the recognition model more comprehensive, and further improving the accuracy of the target information.

[0141] Further, in the training process of the recognition model, since the length of each text is not the same, and since the MiniBert model used cannot cover all the texts due to the high dimension provided by BERT, a part of the text with too long length is cut off. And due to the limitation of the computing capacity of the Graphics Processing Unit (GPU), a larger scale model cannot be used to predict the test data. Therefore, the data features saved in the 40th cycle are used, and the last fully connected layer of the model is removed to obtain a new feature dimension (for example, 1024 dimensions) as new pre-training data, and the transfer learning training is performed to reduce the difference between the source domain and the target domain and improve the accuracy.

[0142] In order to further demonstrate the advantages of the recognition model provided by the embodiments of the present disclosure, some data is provided as follows.

[0143] Figure 7A The AUC curve diagram corresponding to the training data set and the test data set according to the embodiments of the present disclosure is schematically shown. Figure 7B The curve diagram of the loss function according to the embodiments of the present disclosure is schematically shown.

[0144] As shown in Figure 7A and 7B The accuracy of the recognition model provided by the embodiments of the present disclosure is about 10% higher than that of the five MMD models, reaching 50.39%. Similarly, it can be seen from the accuracy curve that overfitting has an impact on the accuracy of the model, and therefore, the model with an epoch of 40 is selected as the final model in the embodiments of the present disclosure, which is used as a way to predict the test set.

[0145] Figure 8 The confusion matrix diagram of the results on the labeled test set in the STN model for the validation set with an epoch of 40 according to the embodiments of the present disclosure is schematically shown.

[0146] As shown in Figure 8As shown, for example, model training is performed on the Guangfa Turing Jupyter platform, and the prediction effect on the validation set is optimized. The model is verified and evaluated through indicators such as Accuracy, N-dimensional confusion matrix, and weighted-F1-score, and the best recognition model result is selected as the final recognition model of the customer call information service recognition. By taking the F1 score of each category and taking the average, the F1 score of this model for the validation set is 0.5055.

[0147] Embodiments of the present disclosure also label the data of the test set.

[0148] Figure 9 An annotation result diagram of labeling the data of the test set according to an embodiment of the present disclosure is schematically shown.

[0149] As Figure 9 shown, account and debit card consultation is the most, accounting for about 65%, most of which is that customers have doubts about transfer or account, and then the customer service manager can process it. The least is the fund-related consultation, which does not appear in the prediction result of the present embodiment. The inaccurate result may be due to the fact that there is less fund-related data in the customer voice, resulting in data imbalance. It is worth mentioning that most of the customer voice data in the training set is account and debit card consultation. Since the present model has proved its convergence and certain accuracy, if the training set label of the present embodiment is replaced by artificially labeled call information, the accuracy will be further improved.

[0150] To sum up, the recognition method of the present embodiment can accurately and comprehensively recognize the target information in the text data. Applying the recognition method to the banking field can accurately serve the customers, thereby improving the service efficiency and service quality of the call service, and improving the customer experience.

[0151] Based on Figures 2-9 the information recognition method, the present embodiment also provides an information recognition device. Based on Figure 1 the scenario described below, the information recognition device of the present embodiment will be described. Figures 10-12 The information recognition device of the present embodiment will be described.

[0152] Figure 10 A block diagram of the information recognition device according to an embodiment of the present disclosure is schematically shown.

[0153] As Figure 10 shown, the information recognition device 1000 can include an acquisition module 1010, a preprocessing module 1020, and a recognition module 1030.

[0154] The obtaining module 1010 is configured to obtain text data to be recognized. The obtaining module 1010 can be configured to perform operation S201 described above, and details are not described herein again.

[0155] The preprocessing module 1020 is configured to pre-process the text data to be recognized to obtain first text features. The preprocessing module 1020 can be configured to perform operation S202 described above, and details are not described herein again.

[0156] The recognition module 1030 is configured to input the first text features into a trained recognition model to recognize target information included in the text data to be recognized. The recognition model is obtained by training a spatial transformation network model based on domain adaptation and using heterogeneous transfer learning, and recognition labels of the recognition model are configured according to business attribute information. The recognition module 1030 can be configured to perform operation S203 described above, and details are not described herein again.

[0157] Figure 11 A block diagram of an information recognition apparatus according to another embodiment of the present disclosure is schematically shown.

[0158] As shown in FIG. 10, the information recognition apparatus 1000 may, for example, further include a filtering module 1040. Figure 11

[0159] The filtering module 1040 is configured to sequentially perform dimension reduction and dimension increase on the first text features to filter out invalid information in the first text features to obtain second text features. The filtering module 1040 can be configured to perform operation S301 described above, and details are not described herein again.

[0160] The recognition module 1030 is further configured to input the first text features into the trained recognition model to recognize the target information included in the text data to be recognized. The recognition module 1030 can be further configured to perform operation S302 described above, and details are not described herein again.

[0161] Figure 12 A block diagram of an information recognition apparatus according to still another embodiment of the present disclosure is schematically shown.

[0162] As shown in FIG. 11, the information recognition apparatus 1000 may, for example, further include a training module 1050. Figure 12

[0163] ​​The training module 1050 is configured to train the recognition model, including: obtaining historical text data. Based on the historical text data, a training data set and a test data set are determined. The training data set is labeled according to a recognition label, and part of the test data of the test data set is labeled according to the recognition label. A statistical distance between a source domain where the training data set is located and a target domain where the test data set is located is taken as a loss function, the labeled training data set and the labeled part of the test data are input into a spatial transformation network model to perform heterogeneous transfer learning, and a trained recognition model is obtained. The training module 1050 can be configured to perform the operations S501-S504 described above, and details are not described herein again.

[0164] Any one or more of the modules, sub-modules, units, sub-units according to the embodiments of the present disclosure, or at least part of the functions of any one or more of the modules, sub-modules, units, sub-units can be implemented in one module. Any one or more of the modules, sub-modules, units, sub-units according to the embodiments of the present disclosure can be split into multiple modules for implementation. Any one or more of the modules, sub-modules, units, sub-units according to the embodiments of the present disclosure can be at least partially implemented as a hardware circuit, such as a field programmable gate array (FPGA), a programmable logic array (PLA), a system on chip, a system on substrate, a system on package, an application specific integrated circuit (ASIC), or any other reasonable way of hardware or firmware through integration or packaging of circuits, or in any one of software, hardware and firmware or in an appropriate combination of any of them. Alternatively, one or more of the modules, sub-modules, units, sub-units according to the embodiments of the present disclosure can be at least partially implemented as computer program modules that can perform corresponding functions when the computer program modules are run.

[0165] For example, any of the acquisition module 1010, the preprocessing module 1020, the identification module 1030, the filtering module 1040, and the training module 1050 can be combined in one module / unit / subunit for implementation, or any of the modules / units / subunits can be split into multiple modules / units / subunits. Alternatively, at least part of the functions of one or more of the modules / units / subunits can be combined with at least part of the functions of other modules / units / subunits and implemented in one module / unit / subunit. According to an embodiment of the present disclosure, at least one of the acquisition module 1010, the preprocessing module 1020, the identification module 1030, the filtering module 1040, and the training module 1050 can be at least partially implemented as a hardware circuit, such as a field programmable gate array (FPGA), a programmable logic array (PLA), a system on chip, a system on substrate, a system on package, an application specific integrated circuit (ASIC), or any other reasonable manner of integrating or packaging a circuit, etc. hardware or firmware, or in any one of software, hardware, and firmware implementation or in a proper combination of any of them. Alternatively, at least one of the acquisition module 1010, the preprocessing module 1020, the identification module 1030, the filtering module 1040, and the training module 1050 can be at least partially implemented as a computer program module that can perform corresponding functions when the computer program module is run.

[0166] It should be noted that the information identification device part in the embodiments of the present disclosure corresponds to the information identification method part in the embodiments of the present disclosure, and the specific implementation details and the resulting technical effects are the same, which will not be repeated here.

[0167] Figure 13 A block diagram of an electronic device suitable for implementing the above-described method according to an embodiment of the present disclosure is schematically shown. Figure 13 The electronic device shown is merely an example and should not bring any limitation to the functions and use range of the embodiments of the present disclosure.

[0168] As Figure 13As shown, the electronic device 1300 according to embodiments of the present disclosure includes a processor 1301 that can perform various appropriate actions and processes according to programs stored in a read only memory (ROM) 1302 or loaded from a storage section 1308 into a random access memory (RAM) 1303. The processor 1301 can include, for example, a general purpose microprocessor (e.g., a CPU), an instruction set processor, and / or a related chip set, and / or a dedicated microprocessor (e.g., an application specific integrated circuit (ASIC)), and so on. The processor 1301 can also include an on-board memory for cache use. The processor 1301 can include a single processing unit or multiple processing units for executing different actions of the method processes according to embodiments of the present disclosure.

[0169] In the RAM 1303, various programs and data required for the operation of the electronic device 1300 are stored. The processor 1301, the ROM 1302, and the RAM 1303 are connected to each other via a bus 1304. The processor 1301 performs various operations of the method processes according to embodiments of the present disclosure by executing programs in the ROM 1302 and / or the RAM 1303. Note that the programs can also be stored in one or more memories other than the ROM 1302 and the RAM 1303. The processor 1301 can also perform various operations of the method processes according to embodiments of the present disclosure by executing programs stored in the one or more memories.

[0170] According to embodiments of the present disclosure, the electronic device 1300 can also include an input / output (I / O) interface 1305, which is also connected to the bus 1304. The electronic device 1300 can also include one or more of the following components connected to the I / O interface 1305: an input section 1306 including a keyboard, a mouse, etc.; an output section 1307 including a display such as a cathode ray tube (CRT), a liquid crystal display (LCD), etc., and a speaker, etc.; a storage section 1308 including a hard disk, etc.; and a communication section 1309 including a network interface card such as a LAN card, a modem, etc. The communication section 1309 performs communication processing via a network such as the Internet. A drive 1310 is also connected to the I / O interface 1305 as necessary. A removable medium 1311 such as a magnetic disk, a magneto-optical disk, a semiconductor memory, etc. is attached to the drive 1310 as necessary, so that a computer program read out therefrom is installed into the storage section 1308 as necessary.

[0171] According to an embodiment of the present disclosure, a method flow according to an embodiment of the present disclosure can be implemented as a computer software program. For example, an embodiment of the present disclosure includes a computer program product comprising a computer program tangibly embodied on a computer readable storage medium, the computer program containing program code for executing the methods illustrated in the flowcharts. In such an embodiment, the computer program can be downloaded and installed from a network by the communication part 1309, and / or installed from the detachable medium 1311. When the computer program is executed by the processor 1301, the above-described functions defined in the system of an embodiment of the present disclosure are executed. According to an embodiment of the present disclosure, the system, device, apparatus, module, unit, and the like described above can be implemented by computer program modules.

[0172] The present disclosure also provides a computer readable storage medium, which can be included in the device / apparatus / system described in the above embodiments, or can exist separately without being assembled into the device / apparatus / system. The above computer readable storage medium tangibly carries one or more programs that, when executed, implement the methods according to an embodiment of the present disclosure.

[0173] According to an embodiment of the present disclosure, the computer readable storage medium can be a non-volatile computer readable storage medium. For example, it can include, but is not limited to, a portable computer diskette, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any appropriate combination thereof. In the present disclosure, the computer readable storage medium can be any tangible medium that contains or stores a program that can be used by or in connection with an instruction execution system, apparatus, or device.

[0174] For example, according to an embodiment of the present disclosure, the computer readable storage medium can include one or more memories of the above-described ROM 1302 and / or RAM 1303 and / or one or more memories other than the ROM 1302 and RAM 1303.

[0175] The computer program product of the second aspect of the disclosure can include a computer readable storage medium. The computer readable storage medium can include instructions. The instructions can include one or both of: instructions for causing a computer to implement a method as described above; and instructions for causing a computer to operate based on a system as described above.

Claims

1. An information identification method, comprising: obtaining text data to be identified; preprocessing the text data to be identified to obtain first text features; the preprocessing the text data to be identified to obtain first text features comprises: determining the length of each text in the text data; comparing the length of each text with a preset length threshold to determine texts with a length greater than the preset length threshold; and cutting a portion from the texts with a length greater than the preset length threshold to obtain the first text features of a preset dimension; the preset length threshold is determined according to the computing capacity of an identification model; after dimension reduction and dimension increase are sequentially performed on the first text features and invalid information in the first text features is filtered, the first text features are input into a trained identification model to identify target information included in the text data to be identified, wherein the identification model is obtained by training a spatial transformation network model based on domain adaptation and using heterogeneous transfer learning, and identification labels of the identification model are configured according to business attribute information; the information identification method further comprises training the identification model, including: obtaining historical text data; determining a training data set and a test data set based on the historical text data, including: down-sampling a portion of the historical text data to obtain the training data set, and randomly sampling another portion of the historical text data to obtain the test data set; labeling the training data set according to the identification labels, and labeling part of the test data of the test data set according to the identification labels; taking a statistical distance between a source domain where the training data set is located and a target domain where the test data set is located as a loss function, inputting the labeled training data set and the labeled part of the test data into the spatial transformation network model for heterogeneous transfer learning to obtain the trained identification model.

2. The information recognition method according to claim 1, wherein Before the labeled training data set and the labeled part of the test data are input into the spatial transformation network model for heterogeneous transfer learning, the method further comprises: mapping the training data set and the test data set to the same dimension.

3. The information recognition method according to claim 2, wherein mapping the training data set and the test data set to a reproducing kernel Hilbert space to make the training data set and the test data set in the same dimension.

4. The information recognition method according to Claim 1, wherein The identification labels include: an account inquiry label, a bank and outlet information inquiry label, an account and debit card label, a transfer and remittance label, a personal credit label, a deposit label, a financial product label, a fund label, a precious metal label, a personal mobile phone bank label, a personal online bank label, a self-service machine label, a messenger label, and a comprehensive label, wherein the comprehensive label is used to represent labels corresponding to business attribute information other than account inquiry, bank and outlet information inquiry, account and debit card, transfer and remittance, personal credit, deposit, financial product, fund, precious metal, personal mobile phone bank, personal online bank, self-service machine, and messenger.

5. An information identification device, comprising: an obtaining module configured to obtain text data to be identified; The preprocessing module is configured to preprocess the text data to be recognized to obtain first text features. The preprocessing module is further configured to determine the length of each text in the text data, compare the length of each text with a preset length threshold, determine texts with a length greater than the preset length threshold, and cut a portion of the texts with the length greater than the preset length threshold to obtain the first text features of a preset dimension. The preset length threshold is determined according to the computing capability of the recognition model. The recognition module is configured to sequentially perform dimension reduction and dimension increase on the first text features, input the first text features into a trained recognition model after filtering invalid information in the first text features, and recognize target information included in the text data to be recognized, wherein the recognition model is obtained by training a spatial transformation network model in a manner of domain adaptation and heterogeneous transfer learning, and recognition labels of the recognition model are configured according to business attribute information. The training module is configured to train the recognition model, including: obtaining historical text data; determining a training data set and a test data set based on the historical text data, including: performing down-sampling on a portion of the historical text data to obtain the training data set, and performing random sampling on another portion of the historical text data to obtain the test data set; labeling the training data set according to the recognition labels, and labeling part of the test data of the test data set according to the recognition labels; taking a statistical distance between a source domain where the training data set is located and a target domain where the test data set is located as a loss function, inputting the labeled training data set and the labeled part of the test data into the spatial transformation network model for heterogeneous transfer learning, and obtaining the trained recognition model. 6.An electronic device, comprising: one or more processors; storage configured to store one or more programs, wherein the one or more programs, when executed by the one or more processors, cause the one or more processors to perform the method according to any one of claims 1-4. 7.A computer-readable storage medium having stored thereon executable instructions that, when executed by a processor, cause the processor to perform the method according to any one of claims 1-4. 8.A computer program product comprising a computer program that, when executed by a processor, implements the method according to any one of claims 1-4.

Images