Fund user classification method and device

Abstract

The present application discloses a fund user classification method and device, which respectively adds a padding sequence of a preset length to the beginning and end of a word sequence obtained by segmenting text information of a fund user to be classified, extracts the semantic features of the padded sequence, and then inputs the obtained word embedding matrix into a trained classification model, and finally determines the classification result of the fund user to be classified by the feature information obtained by the convolution operation of the classification model, wherein the classification model is trained with training texts marked with category labels. By adding padding sequences to the beginning and end of a word sequence, the convolution search space of the classification model is expanded, and since the text boundary information is no longer at the edge of the search space, the classification model can extract more boundary feature information when extracting feature information, thereby realizing targeted processing of text boundary information, obtaining stronger text boundary features, and thus improving the accuracy of the classification results.

Description

Technical Field

[0001] The present application relates to the technical field of fund user classification processing, and more specifically, to a fund user classification method and device. Background Art

[0002] With the advancement of science and technology, people's work and life have entered the digital age. Various applications need to obtain data during operation and also generate data. There are many types of data, and sometimes in order to improve work efficiency, data needs to be classified. Taking fund companies as an example, newly registered institutional users of fund users need to be labeled with multiple categories of labels due to the compliance requirements of multiple regulatory departments and in order to improve the efficiency of fund user information management. For example, it is necessary to determine the customer type of fund users, the source of funds type, the type of public offering source investment, etc., that is, it is necessary to classify fund users according to their types.

[0003] Among the existing classification methods, there is a method in which a model is used to classify funds according to certain classification standards based on the nature information of the fund users to be classified. Although this method can achieve high-efficiency classification, the accuracy of the classification result is relatively poor. Summary of the invention

[0004] In view of the above problems, this application provides a fund user classification method and device to solve the problem of poor classification accuracy of existing classification methods. The specific solution is as follows:

[0005] In a first aspect, a fund user classification method is provided, comprising:

[0006] Acquire text information of the fund user to be classified, wherein the text information includes attribute information of the fund user;

[0007] Performing word segmentation processing on the text information to obtain a word sequence;

[0008] Adding a padding sequence of a preset length to the beginning and the end of the word sequence to obtain a padded sequence;

[0009] Extracting semantic features from the padded sequence to obtain a word embedding matrix;

[0010] The word embedding matrix is ​​input into the trained classification model so that the classification model can perform convolution operation on the word embedding matrix to obtain feature information, and predict the classification result of the fund user to be classified based on the feature information, wherein the classification model is trained with training text marked with category labels.

[0011] In a second aspect, a fund user classification device is provided, comprising:

[0012] A text information acquisition unit, used to acquire text information of the fund user to be classified, wherein the text information includes attribute information of the fund user;

[0013] A word sequence acquisition unit, used for performing word segmentation processing on the text information to obtain a word sequence;

[0014] A padded sequence acquisition unit, used to add a padded sequence of a preset length to the beginning and the end of the word sequence to obtain a padded sequence;

[0015] A semantic feature extraction unit, used to extract semantic features from the padded sequence to obtain a word embedding matrix;

[0016] A classification result determination unit is used to input the word embedding matrix into a trained classification model so that the classification model can perform a convolution operation on the word embedding matrix to obtain feature information, and predict the classification result of the fund user to be classified based on the feature information, wherein the classification model is trained with training text marked with category labels.

[0017] By means of the above technical scheme, the present application performs word segmentation processing on the text information of the fund users to be classified, obtains a word sequence, and adds a padding sequence of a preset length at the beginning and end of the word sequence to obtain a padded sequence, performs semantic feature extraction on the padded sequence to obtain a word embedding matrix, inputs the word embedding matrix into the trained classification model, and the classification model performs convolution operation on the word embedding matrix to obtain feature information, and determines the classification result of the fund users to be classified based on the feature information, wherein the classification model is obtained by training the training text marked with category labels. Since the padding sequence is added at the beginning and end of the word segmentation processing result of the text information, the search space of the classification model when performing the convolution operation is expanded, so that the boundary information of the text is no longer located at the edge of the search space, so when the classification model extracts feature information through the convolution operation, it can extract more boundary feature information, realizes the targeted processing of the text boundary information, obtains stronger text boundary features, and thus improves the accuracy of the classification result. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Various other advantages and benefits will become apparent to those of ordinary skill in the art by reading the detailed description of the preferred embodiments below. The accompanying drawings are only for the purpose of illustrating the preferred embodiments and are not to be considered as limiting the present application. Also, the same reference symbols are used throughout the accompanying drawings to represent the same components. In the accompanying drawings:

[0019] Figure 1 is a flow chart of a fund user classification method according to an exemplary embodiment;

[0020] Figure 2is a schematic diagram of a process of obtaining a word embedding matrix according to an exemplary embodiment;

[0021] Figure 3 is a schematic diagram of a process of performing a convolution operation according to an exemplary embodiment;

[0022] Figure 4 is a schematic diagram of a process for determining a classification result according to an exemplary embodiment;

[0023] Figure 5 is a flow chart of a fund user classification method according to another exemplary embodiment;

[0024] Figure 6 is a schematic diagram of a process of performing a convolution operation according to another exemplary embodiment;

[0025] Figure 7 is a flow chart of a fund user classification method according to another exemplary embodiment;

[0026] Figure 8 The figure is a schematic diagram of the structure of a fund user classification device according to an exemplary embodiment. DETAILED DESCRIPTION

[0027] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0028] The present application provides a fund user classification scheme that can complete the fund user classification task, which can be a task of determining the classification result of the base user based on the attribute information of the fund user to be classified, wherein the classification result can be one or more types of category labels. It should be noted that there is no direct relationship between the various types of labels, and they correspond to different classification standards. For example, based on the attribute information of the fund user, the customer type, internal transaction type, source of funds type, public offering source investment type, importance type, etc. of the fund user can be determined.

[0029] The present application solution can be implemented based on a terminal with data processing capabilities, which can be a mobile phone, computer, server, cloud, etc.

[0030] Figure 1 is a flow chart showing a method for classifying fund users according to an exemplary embodiment. Figure 1 As shown, a fund user classification method may include the following steps:

[0031] Step S101: Obtain text information of fund users to be classified.

[0032] Among them, the fund users to be classified are the objects that need to be classified. The fund users to be classified have certain attribute information, which is used to characterize the nature and characteristics of the fund users. These attribute information can be expressed in the form of text information, which is the basis for classifying the fund users; the text information can be expressed as a sentence or a combination of multiple sentences, and can be text information that is manually input and available for acquisition.

[0033] Step S102: perform word segmentation processing on the text information to obtain a word sequence.

[0034] Specifically, word segmentation can be to divide text information into units according to certain rules, where a unit can be a word or a punctuation mark; each unit corresponds to each element that constitutes a word sequence, and the word sequence can be obtained by arranging each unit according to the order in which it appears in the text information. Take the text information "CCB Fund CCB Branch Asset Management Plan" as an example, Figure 2 A possible schematic diagram of the word segmentation process is shown.

[0035] Optionally, after the text information is segmented, the method may further include: performing a stop word removal process on the word sequence obtained by the word segmentation process to obtain a final word sequence, wherein the stop word may be a punctuation mark or other word that has no practical significance for classification. By performing the stop word removal process, meaningless content in the word sequence can be filtered out, thereby reducing the amount of information used for classification and reducing resource consumption.

[0036] Step S103: adding a padding sequence of a preset length to the beginning and the end of the word sequence to obtain a padded sequence.

[0037] Among them, the filling sequence is a sequence of preset specific content. By extracting semantic features from the filling sequence, the corresponding specific content can be obtained; the composition of the filled sequence can include, in order, the filling sequence, the word sequence and the filling sequence. Take the word sequence obtained by word segmentation of "CCB Fund CCB Branch Asset Management Plan" as an example, Figure 2 A schematic diagram showing a possible composition of a filled sequence.

[0038] Step S104: extract semantic features from the padded sequence to obtain a word embedding matrix.

[0039] Specifically, the process of performing semantic feature extraction on the padded sequence to obtain a word embedding matrix may include: performing semantic feature extraction on each element constituting the padded sequence one by one to obtain a semantic feature extraction result corresponding to the element, wherein the result may be represented in the form of a word vector, and arranging each word vector according to the order of the corresponding element in the padded sequence to obtain a word embedding matrix.

[0040] For example, take the text information "CCB Fund Construction Bank Branch Asset Management Plan" as an example. Figure 2 A possible composition diagram of a word embedding matrix is ​​shown, Figure 2 As shown in the figure, when the length of the padded sequence is 15, that is, the padded sequence includes 15 elements, and the dimension of the generated word vector is 5, the resulting word embedding matrix is ​​15×5 in size, where each row corresponds to a word vector.

[0041] The method for extracting semantic features may be to use the neural network model word2vec to extract semantic features. Word2vec can represent a word in natural language as a word vector of uniform dimension. The word vector generated by word2vec can better express the similarity and analogy between different words. In addition, any one of the methods such as bag of words model, topic model, or word embedding may be used to extract semantic features.

[0042] Step S105: input the word embedding matrix into the trained classification model to determine the classification result of the fund user to be classified.

[0043] Specifically, after the word embedding matrix is ​​input into the trained classification model, the classification model performs a convolution operation on the word embedding matrix to obtain feature information, and predicts the classification result of the fund user to be classified based on the obtained feature information. Figure 3 1 is a schematic diagram of a possible process of performing a convolution operation according to an exemplary embodiment. Taking the text information of the fund user to be classified as "CCB Fund CCB Branch Asset Management Plan" as an example, steps S102-S103 are executed on the above text information to obtain a padded sequence, wherein the length of the padded sequence used is 4, and the convolution kernel with a size of 5 (such as Figure 3 As shown in the box in the figure, the word embedding matrix corresponding to the padded sequence is convolved from the beginning to the end of the padded sequence according to the convolution step size of 1, combined with Figure 3As shown, the convolution kernel can extract 4 pieces of left boundary feature information containing partial padding sequences and the beginning “Jianxin” of the word sequence, 3 pieces of internal feature information containing only the word sequence, and 4 pieces of right boundary feature information containing partial padding sequences and the end “plan” of the word sequence. By adding padding sequences to the word sequence, the convolution search space is expanded, and more boundary feature information is extracted, thereby improving the accuracy of the classification result. It should be noted that the object of the convolution operation of the classification model is the word embedding matrix, and the word embedding matrix is obtained by extracting semantic features from the word sequence. There is a corresponding relationship between the two, so the word sequence can be used as a simplified representation of the word embedding matrix. That is to say, Figure 3 a word in

[0044] is actually a word vector. The classification model is trained with training texts labeled with class labels. Specifically, the n-fold cross-validation mode can be used to train the classification model. Exemplarily, the training dataset includes training texts and corresponding class labels. When training the classification model, the training dataset can be randomly divided into n parts, and one of them is used as the validation set in turn, while the remaining several datasets are used as the training set to train and optimize the model parameters, thereby obtaining the trained classification model.

[0045] The above solution exemplarily gives a classification model based on a convolutional neural network. In addition, the classification model can also adopt other machine learning models, such as: support vector machine, decision tree, nearest neighbor node algorithm, random forest, gradient boosting decision tree, multi-layer perceptron, deep neural network, recurrent neural network and other models.

[0046] The fund user classification method provided by the embodiments of this application performs word segmentation on the text information of the fund user to be classified to obtain a word sequence, adds padding sequences with a preset length at the beginning and end of the word sequence respectively to obtain a padded sequence, performs semantic feature extraction on the padded sequence to obtain a word embedding matrix, inputs the word embedding matrix into the trained classification model, the classification model performs a convolution operation on the word embedding matrix to obtain feature information, and determines the classification result of the fund user to be classified according to the feature information, where the classification model is trained with training texts labeled with class labels. Since padding sequences are added at the beginning and end of the word segmentation result of the text information, the search space for the convolution operation of the classification model is expanded, so that the boundary information of the text is no longer at the edge of the search space. Therefore, when the classification model extracts feature information through convolution operation, more boundary feature information can be extracted, realizing the targeted processing of the text boundary information, obtaining stronger text boundary features, and further improving the accuracy of the classification result.

[0047] In some embodiments provided in the present application, the padding sequence may be an empty sequence, wherein the empty sequence is a sequence that does not contain any elements, and on this basis, by performing semantic feature extraction on an element constituting the padding sequence, a zero vector of a preset dimension may be obtained, that is, by performing semantic feature extraction on a padding sequence of a preset length, a zero matrix with a preset length of rows and a preset dimension of columns may be obtained.

[0048] In some embodiments provided in the present application, the classification model includes a convolution layer. In order to extract rich feature information of multi-element relationships between words, the convolution layer may include multiple convolution kernels, wherein the convolution kernels may have multiple different sizes, and there may be multiple convolution kernels of each size. For example, the size of the convolution kernel may be set to 2, 3, 4, and 5, so as to extract feature information representing binary, ternary, quaternary, and quinary multi-element relationships between words.

[0049] It should be noted that the following width and height are only exemplary definitions for the convenience of explanation. Specifically, the width of the convolution kernel can be the dimension of the word vector. On this basis, the size difference of the convolution kernel is reflected in the height difference of the convolution kernel. The reason is that the object of the convolution kernel convolution is the word embedding matrix composed of word vectors, and the word vector is the smallest unit that represents the semantic features of a word. In order to obtain complete and useful semantic feature information of a word, the width of the convolution kernel can be set to the dimension of the word vector. If it is smaller, the complete semantic feature information cannot be obtained. If it is larger, useless information is obtained, which wastes resources. Therefore, the width of the convolution kernel is fixed and equal to the dimension of the word vector. On the basis of the above, the height of the convolution kernel corresponds to the moving direction of the convolution kernel, that is, the height of the convolution kernel corresponds to the arrangement direction of the word vectors corresponding to each word in the word embedding matrix. Exemplarily, if a row in the word embedding matrix corresponds to a word vector, that is, each word vector is arranged vertically in the word embedding matrix, then the convolution kernel moves in the vertical direction, and the length of the convolution kernel in the vertical direction is regarded as the height of the convolution kernel.

[0050] In a possible implementation, the preset length may be the maximum size of each convolution kernel minus 1, and the maximum size is not less than 2. Exemplarily, if the size of the convolution kernel is 2, 3, 4, and 5, the preset length is 4.

[0051] According to the above rules, the preset length is set, which not only expands the convolution search space, makes it possible to extract effective feature information, especially the extraction of boundary feature information, but also reduces the resources occupied during classification. If the preset length is long, useless feature information will be extracted, that is, feature information that only contains elements in the filling sequence but not in the word sequence, which occupies more storage and computing resources and causes resource waste; if the preset length is short, the elements at the boundary of the text information are not fully utilized, which may be manifested in that a few convolutions are performed on the boundary elements of the text information, while multiple convolutions are performed on the middle elements of the text information, and sufficient boundary feature information cannot be extracted.

[0052] Figure 4 FIG. 1 is a schematic diagram of a process for determining a classification result according to an exemplary embodiment, showing a possible composition of a word embedding matrix, and a schematic diagram of a process for a classification model to process the word embedding matrix to determine a classification result. Figure 4 As shown in the figure, the convolution kernel size of the classification model can be 2, 3, 4 and 5, then the padding sequence can be an empty sequence of length 4. The classification model performs a convolution operation on the word embedding matrix to extract feature information and processes the feature information, such as using a pooling function on the feature information obtained by the convolution kernel of the same size to obtain the corresponding pooling results, concatenating the pooling results to obtain the final feature vector representation, and obtaining the probability of belonging to each category by the fully connected layer, thereby determining the final category label.

[0053] In the embodiment provided in the present application, the classification model includes an input layer. In order to solve the problem that the length of the text information of different fund users to be classified is different, the length of the input layer is set, and the length of the input layer can be a preset fixed length. Although the input layer of the classification model is a word embedding matrix, the length of the word embedding matrix is ​​associated with the length of the text information of the fund user to be classified (assuming that the width of the word embedding matrix is ​​consistent with the dimension of the word vector), which is specifically manifested as follows: the longer the text information, the longer the word sequence obtained by word segmentation, and accordingly, the length of the padded sequence and the length of the corresponding word embedding matrix are also larger.

[0054] On the basis of the above, before step S104, extracting semantic features from the padded sequence, the following may also be included:

[0055] The length of the padded sequence is corrected to match the length of the input layer, thereby obtaining a corrected padded sequence.

[0056] In a possible implementation, the process of correcting the length of the padded sequence to match the length of the input layer may include:

[0057] If the length of the padded sequence is greater than the length of the input layer, starting from the beginning of the padded sequence, a sequence with a length equal to the length of the input layer is intercepted, and the intercepted sequence is used as the modified padded sequence;

[0058] If the length of the padded sequence is less than the length of the input layer, a padded sequence having a length equal to the difference between the length of the input layer and the length of the padded sequence is added to the end of the padded sequence.

[0059] On the basis of the above, step S104, extracting semantic features from the padded sequence, may include:

[0060] Semantic feature extraction is performed on the corrected padded sequence.

[0061] The reason for setting the length of the input layer of the classification model to a fixed value is that the classification model needs to use a fully connected layer to determine the probability that the fund user to be classified belongs to each category label. If the length of the input layer is not fixed, the number of weight parameters of the fully connected layer is also not fixed, which will cause dynamic changes in the network and fail to achieve the purpose of parameter training.

[0062] When setting the length of the input layer of the classification model, the length of the training text can be mainly considered, so that most of the training text can be used when training the model to obtain more accurate model parameters. Therefore, by reasonably setting the length of the input layer of the classification model, unnecessary resource consumption can be reduced, a more accurate classification model can be obtained through training, and the accuracy of the classification model in classification prediction can be improved.

[0063] Figure 5 is a flow chart showing another fund user classification method according to an exemplary embodiment. Figure 5 As shown, a fund user classification method may include the following steps:

[0064] Step S201: Obtain text information of fund users to be classified.

[0065] The text information includes multiple text information of different types, each of which is manually input for classification, and each text information corresponds to a type of attribute information of the fund user. Exemplarily, the text information may include: Fuzhou Electric Power Material Co., Ltd.; data source is the "Marketing" department; company type is "non-financial enterprise" type; parent company type is "state-owned enterprise" type, etc. The above text information points to different types, enriching the text information used for classification.

[0066] Step S202: perform word segmentation processing on each piece of text information to obtain a corresponding word sequence.

[0067] The process of performing word segmentation on each piece of text information to obtain a corresponding word sequence is consistent with the above-mentioned step S102 and will not be repeated here.

[0068] Step S203 , concatenate the word sequences, and add a filling sequence of a preset length to the beginning, the end, and between two adjacent word sequences of the concatenated word sequence to obtain a filled sequence.

[0069] Steps S204-S205 are consistent with the above-mentioned steps S104-S105 and will not be repeated here.

[0070] Figure 6 is a schematic diagram of a process of performing a convolution operation according to another exemplary embodiment, wherein there are two pieces of text information of fund users to be classified, namely, "CCB Fund Construction Bank Branch Asset Management Plan" and "Internet Finance". Figure 6 The padded sequence obtained by executing steps S202-S203 on the above text information is shown. The padded sequence consists of two corresponding word sequences and three padded sequences of length 4. The padded sequences are distributed between the two corresponding word sequences and at the beginning and end of the word sequence obtained by splicing the two word sequences. The convolution kernel of size 5 performs convolution operation on the word embedding matrix corresponding to the padded sequence according to the direction of the arrow in the figure. The specific convolution operation process is consistent with that described above and will not be repeated here. It should be noted that the object of the classification model convolution operation is the word embedding matrix obtained by extracting semantic features from the padded sequence. Figure 6 It is just a simplified diagram of the process based on the correspondence between the padded sequence and the word embedding matrix.

[0071] The above-mentioned embodiment provides a possible way of text information fusion for the case where the text information types are diverse. By splicing the word sequences corresponding to different types of text information, and adding a padding sequence of preset length at the beginning, end and between each two adjacent word sequences of the spliced ​​word sequence, it not only realizes the fusion of different types of text information, but also realizes the separation of different types of text information. The padded sequence containing various types of text information is used as the basis for the classification model to perform convolution operations, making full use of different types of text information. Moreover, when the length of the above-mentioned padding sequence is set reasonably, effective separation of different types of text information can be achieved. Exemplarily, if the length of the padding sequence is set to at least the size of the maximum convolution kernel of the classification model minus one, each feature information extracted by the convolution operation corresponds only to one piece of text information, and feature information containing the tail end of the previous text information and the head end of the next text information in the adjacent text information will not be extracted, thereby realizing the separation of different types of text information.

[0072] Figure 7is a flow chart showing another fund user classification method according to an exemplary embodiment. Figure 7 As shown, a fund user classification method may include the following steps:

[0073] Steps S301-S302 are consistent with the above-mentioned steps S201-S202 and will not be repeated here.

[0074] Step S303: For each word sequence, a padding sequence of a preset length is added to the beginning and the end of the word sequence to obtain a padded sequence corresponding to the word sequence.

[0075] Step S304: merge the filled sequences corresponding to the word sequences to obtain the final filled sequence.

[0076] In a possible implementation, the filled sequence corresponding to the fused word sequences may include:

[0077] The padded sequences corresponding to each word sequence are concatenated, and the length of the padded sequence between each two adjacent word sequences is adjusted to a preset length.

[0078] Steps S305-S306 are consistent with the above-mentioned steps S204-S205 and will not be repeated here.

[0079] The above-mentioned embodiment provides another possible way of text information fusion for the situation where the text information types are diverse. By adding a padding sequence of a certain length at the beginning and end of each word sequence corresponding to different types of text information, the padded sequences corresponding to each word sequence are fused to obtain the final padded sequence, which is used as the object of subsequent semantic feature extraction, thereby realizing the utilization of different types of text information.

[0080] The fund user classification method provided in the embodiment of the present application is suitable for situations where the text information types are diverse. On the basis of ensuring full utilization of text boundary information, multiple text information of different types is used for classification. Accordingly, when the classification model is trained with training texts marked with category labels, the training texts used also contain multiple text information of different types, and the length of the input layer of the classification model also increases accordingly. Therefore, the trained classification model can cope with the situation where the text information is multiple text information of different types, which provides the possibility of adding additional text information for classification without adding new network nodes, thereby avoiding an increase in the complexity of the network structure.

[0081] The fund user classification method provided in the embodiment of the present application can be applied to solve the fund user classification task, which can be to determine a category label of the fund user to be classified, or to determine multiple different types of category labels of the fund user to be classified, and there is no direct relationship between the category labels of each type, that is to say, multiple category labels of a fund user to be classified correspond to different classification standards. In order to complete the multi-category label classification task of the fund user to be classified, the word embedding matrix obtained by the fund user classification method can be input into multiple classification models respectively, and each classification model outputs a classification result, i.e., a category label. It should be noted that the classification model corresponds one-to-one to the classification standard, and each classification model is trained with a training text marked with a category label corresponding to the classification standard.

[0082] The fund user classification method provided in the embodiment of the present application can be applied in a fund company to complete the task of classifying fund users according to their attribute information and determining their category labels, wherein the text information of the fund users to be classified can be the attribute information of the fund users, such as the name of the fund users, etc.

[0083] The fund user classification device provided in the embodiment of the present application is described below. The fund user classification device described below and the fund user classification method described above can be referenced to each other.

[0084] See also Figure 8 , Figure 8 This is a schematic diagram of the structure of a fund user classification device disclosed in an embodiment of the present application.

[0085] like Figure 8 As shown, the device may include:

[0086] A text information acquisition unit 11 is used to acquire text information of the fund user to be classified, wherein the text information includes attribute information of the fund user;

[0087] A word sequence acquisition unit 12 is used to perform word segmentation processing on the text information to obtain a word sequence;

[0088] A padded sequence acquisition unit 13 is used to add a padded sequence of a preset length to the beginning and the end of the word sequence to obtain a padded sequence;

[0089] A semantic feature extraction unit 14, used to extract semantic features from the padded sequence to obtain a word embedding matrix;

[0090] The classification result determination unit 15 is used to input the word embedding matrix into the trained classification model so that the classification model can perform convolution operation on the word embedding matrix to obtain feature information, and predict the classification result of the fund user to be classified based on the feature information, wherein the classification model is trained with training text marked with category labels.

[0091] In a possible implementation, the word sequence acquisition unit 12 performs word segmentation processing on the text information to obtain a word sequence, which may include:

[0092] The text information is segmented, and the word sequence obtained by the word segmentation is processed to remove stop words to obtain a final word sequence, wherein the stop words include punctuation marks.

[0093] In a possible implementation, the process of the text information acquiring unit 11 acquiring the text information of the fund user to be classified may include:

[0094] Acquire multiple pieces of text information of different types of fund users to be classified, each piece of text information corresponding to a type of attribute information of the fund user;

[0095] The word sequence acquisition unit 12 performs word segmentation processing on the text information to obtain a word sequence, which may include:

[0096] Perform word segmentation on each piece of text information to obtain the corresponding word sequence;

[0097] The process in which the padded sequence acquisition unit 13 adds a padded sequence of a preset length to the beginning and the end of the word sequence to obtain the padded sequence may include:

[0098] For each word sequence, a padding sequence of a preset length is added to the beginning and the end of the word sequence to obtain a padded sequence corresponding to the word sequence;

[0099] The filled sequences corresponding to each word sequence are merged to obtain the final filled sequence;

[0100] or,

[0101] Each word sequence is concatenated, and a padding sequence of a preset length is added to the beginning, the end and between each two adjacent word sequences of the concatenated word sequence to obtain a padded sequence.

[0102] In a possible implementation, the process of the filled sequence acquisition unit 13 fusing the filled sequences corresponding to the word sequences may include:

[0103] The padded sequences corresponding to each word sequence are concatenated, and the length of the padded sequence between each two adjacent word sequences is adjusted to a preset length.

[0104] In a possible implementation manner, the filling sequence is an empty sequence.

[0105] In a possible implementation, the classification model includes a convolution layer, and the convolution layer includes a plurality of convolution kernels;

[0106] The preset length is the maximum size of each convolution kernel minus 1, and the maximum size is not less than 2.

[0107] In a possible implementation, the classification model includes an input layer, and the length of the input layer is a preset fixed length;

[0108] The fund user classification device may further include:

[0109] A sequence length correction unit, used to correct the length of the padded sequence to match the length of the input layer, thereby obtaining a corrected padded sequence;

[0110] The process of extracting semantic features from the padded sequence by the semantic feature extraction unit 14 may include:

[0111] Semantic feature extraction is performed on the corrected padded sequence.

[0112] Finally, it should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the presence of other identical elements in the process, method, article or device including the elements.

[0113] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The various embodiments can be combined as needed, and the same or similar parts can refer to each other.

[0114] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A fund user classification method, characterized in that: include: Acquire text information of the fund user to be classified, wherein the text information includes attribute information of the fund user; Performing word segmentation processing on the text information to obtain a word sequence; Adding a padding sequence of a preset length to the beginning and the end of the word sequence to obtain a padded sequence, so as to expand the search space of the classification model when performing a convolution operation, so that the boundary information of the text is no longer located at the edge of the search space; Extracting semantic features from the padded sequence to obtain a word embedding matrix; Inputting the word embedding matrix into the trained classification model so that the classification model performs a convolution operation on the word embedding matrix to obtain feature information, and predicting the classification result of the fund user to be classified according to the feature information, wherein the classification model is trained with training texts marked with category labels; The text information includes multiple pieces of text information of different types, each piece of text information corresponds to a kind of attribute information of the fund user; The word segmentation processing of the text information to obtain a word sequence includes: Perform word segmentation on each piece of text information to obtain the corresponding word sequence; The step of adding a padding sequence of a preset length to the beginning and the end of the word sequence to obtain a padded sequence comprises: For each word sequence, a padding sequence of a preset length is added to the beginning and the end of the word sequence to obtain a padded sequence corresponding to the word sequence; The filled sequences corresponding to each word sequence are merged to obtain the final filled sequence; or, Concatenate each word sequence, and add a padding sequence of a preset length to the beginning, the end, and between two adjacent word sequences of the concatenated word sequence to obtain a padded sequence; The filled sequence corresponding to each word sequence is fused, including: splicing the padded sequences corresponding to each word sequence, and adjusting the length of the padded sequence between each two adjacent word sequences to a preset length; Wherein, the filling sequence is an empty sequence; Wherein, the classification model includes a convolution layer, and the convolution layer includes a plurality of convolution kernels; The preset length is the maximum size of each convolution kernel minus 1, and the maximum size is not less than 2; Among them, in the case of various types of text information, by splicing the word sequences corresponding to different types of text information, and adding padding sequences of preset lengths to the head, tail and between each adjacent word sequence of the spliced ​​word sequence, different types of text information can be fused and separated. The padded sequence containing various types of text information is used as the basis for convolution operations of the classification model, which makes full use of different types of text information. Moreover, when the length of the padding sequence is set reasonably, different types of text information can be effectively separated.

2. The method according to claim 1, characterized in that The classification model includes an input layer, and the length of the input layer is a preset fixed length; Before extracting semantic features from the padded sequence, the method further includes: Correcting the length of the padded sequence to match the length of the input layer to obtain a corrected padded sequence; The extracting semantic features from the padded sequence includes: Semantic feature extraction is performed on the corrected padded sequence.

3. A fund user classification device, characterized in that: include: A text information acquisition unit, used to acquire text information of the fund user to be classified, wherein the text information includes attribute information of the fund user; A word sequence acquisition unit, used for performing word segmentation processing on the text information to obtain a word sequence; A padded sequence acquisition unit is used to add a padded sequence of a preset length to the beginning and the end of the word sequence to obtain a padded sequence, so as to expand the search space when the classification model performs a convolution operation, so that the boundary information of the text is no longer located at the edge of the search space; A semantic feature extraction unit, used to extract semantic features from the padded sequence to obtain a word embedding matrix; A classification result determination unit, used for inputting the word embedding matrix into a trained classification model, so that the classification model performs a convolution operation on the word embedding matrix to obtain feature information, and predicts the classification result of the fund user to be classified according to the feature information, wherein the classification model is trained with training texts marked with category labels; The process of the text information acquisition unit acquiring the text information of the fund user to be classified includes: Acquire multiple pieces of text information of different types of fund users to be classified, each piece of text information corresponding to a type of attribute information of the fund user; The word sequence acquisition unit performs word segmentation processing on the text information to obtain a word sequence, including: Perform word segmentation on each piece of text information to obtain the corresponding word sequence; The process of the padded sequence acquisition unit adding a padded sequence of a preset length to the beginning and the end of the word sequence to obtain the padded sequence includes: For each word sequence, a padding sequence of a preset length is added to the beginning and the end of the word sequence to obtain a padded sequence corresponding to the word sequence; The filled sequences corresponding to each word sequence are merged to obtain the final filled sequence; or, Concatenate each word sequence, and add a padding sequence of a preset length to the beginning, the end, and between two adjacent word sequences of the concatenated word sequence to obtain a padded sequence; The process of the filled sequence acquisition unit fusing the filled sequences corresponding to the word sequences includes: splicing the padded sequences corresponding to each word sequence, and adjusting the length of the padded sequence between each two adjacent word sequences to a preset length; Wherein, the filling sequence is an empty sequence; The classification model includes a convolution layer, and the convolution layer includes a plurality of convolution kernels; the preset length is the maximum size of each convolution kernel minus 1, and the maximum size is not less than 2; Among them, in the case of various types of text information, by splicing the word sequences corresponding to different types of text information, and adding padding sequences of preset lengths to the head, tail and between each adjacent word sequence of the spliced ​​word sequence, different types of text information can be fused and separated. The padded sequence containing various types of text information is used as the basis for convolution operations of the classification model, which makes full use of different types of text information. Moreover, when the length of the padding sequence is set reasonably, different types of text information can be effectively separated.

Images