CN114724167A - Marketing text recognition method and system - Google Patents

Abstract

The invention provides a marketing text recognition method and a marketing text recognition system, which comprise the following steps: acquiring a text to be recognized and preprocessing the text; constructing a text graph of the text to be recognized based on the preprocessed text to be recognized; generating text level word representation based on a text graph of a text to be recognized, and generating text representation by combining the embedded representations of all labels; based on the text representation, a classifier is adopted to obtain a result of whether the text to be recognized belongs to the marketing text; the method for acquiring the embedded representation of the label comprises the following steps: generating probability distribution of the subject words based on the text graph and the labels of the training set, mapping the probability distribution of the subject words to a label vector space, and learning the correlation relation and semantic information among the labels under the guidance of the label graph to obtain the embedded representation of the labels. The purpose of generating complete label embedding is achieved, more information related to classification is captured by jointly learning words and labels, and the marketing text recognition precision is improved.

Description

Marketing text recognition method and system

technical field

The invention belongs to the technical field of natural language processing, and in particular relates to a marketing text recognition method and system.

Background technique

The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

Every day, social media platforms publish a large number of marketing articles with advertising content for promotional purposes. In order to eliminate readers' disgust and win the trust of potential customers, marketing content is often hidden in ordinary article content, which is difficult to identify and discover. Different from traditional media, some self-media editors carefully design advertising content for profit, and even exaggerate facts and edit false information, which will not only mislead consumers and harm consumers' interests, but also damage a healthy network environment. Therefore, corresponding methods and systems are urgently needed to detect self-media content marketing articles.

Most studies treat the detection of content marketing articles as a text classification problem. Among the existing methods, the methods for solving text classification tasks are mainly divided into the following categories: (1) Traditional text classification methods include Naive Bayes, maximum entropy, decision tree, support vector machine, etc. These methods mainly rely on manual annotation. (2) Deep learning methods, such as TextRNN, TextRCNN, and fastText, can automatically obtain classification features, but these methods focus on the locality of words and lack long distance and non-consecutive word interactions; (3) methods based on graph neural networks, such as TextGCN, HyperGAT, TextING, can directly process complex structured data and perform priority mining on global features, but such methods do not consider detailed Granular label information as well as label-related textual information.

In recent years, some studies have found that labels are directly related to text classification and can help models obtain information more relevant to classification. For multi-label classification tasks, a large number of labels are often involved, and the relationship between categories is complex, and it is difficult to find a reasonable way to describe them. Therefore, effectively mining label information is the key to the success of multi-label text classification learning. To exploit label information to improve classification performance, deep extreme multi-label learning method (DXML) captures label dependencies by exploring label structure, sequence generation model (SGM) utilizes long short-term memory network (LSTM, Long Short-Term Memory) to process label sequences Dependencies, capturing complex label relationships. Although they learn label information from different aspects to obtain label relationships, they ignore the role of label co-occurrence information in capturing label relationships, and it is difficult to reasonably describe label relationships to help the model improve the classification effect. The Label Embedding Attention Model (LEAM) combines words and labels to obtain embedded representations of text, and the Explicit Interaction Model (EXAM) utilizes class representations to obtain word-label interaction information. These methods consider the label information in the form of label embedding, but when the label space is vectorized, the label vector lacks the feature information transfer between labels, so that the label representation cannot cover the complete semantics of the label space. Label-specific attention network (LSAN) utilizes a self-attention mechanism to identify label-specific information, and some autoencoder-based methods produce similar text label scores through a ranking-based autoencoder architecture. However, such methods assume that all labels are independent of each other and do not fully consider label semantics and overall label correlation.

SUMMARY OF THE INVENTION

In order to solve the technical problems existing in the above background technology, the present invention provides a marketing text recognition method and system, which achieves the purpose of capturing the global semantics and overall correlation of tags, generating complete tag embedding, and jointly learning words and tags to capture more Classification-related information that improves the accuracy of marketing text recognition.

In order to achieve the above object, the present invention adopts the following technical solutions:

A first aspect of the present invention provides a marketing text recognition method, comprising:

Obtain the text to be recognized and preprocess it;

Based on the preprocessed text to be recognized, construct a text map of the text to be recognized;

Generate text-level word representations based on the text graph of the text to be recognized, and combine the embedded representations of all tags to generate text representations;

Based on the text representation, the classifier is used to obtain the result of whether the text to be identified belongs to the marketing text;

Among them, the method of obtaining the embedded representation of the label is: based on the text map of the training set and its labels, generate the probability distribution of the topic words, map the probability distribution of the topic words to the label vector space, and learn the correlation between the labels under the guidance of the label map. Sexual relationship and semantic information, get the embedded representation of the label.

Further, the preprocessing includes cleaning non-text data, removing stop words, removing low frequency words, removing high frequency words and morphological restoration.

Further, the construction method of the text graph is: for a text, count the co-occurrence times of words in a fixed sliding window, take each word as a vertex of the text graph, and the co-occurrence times between words as an edge of the text graph.

Further, the method for constructing the label graph is: for all labels in the training set, count the co-occurrence times of labels in a fixed sliding window, take each label as a vertex and the label co-occurrence times as an edge.

Further, the generation method of described topic word probability distribution is:

Using the Dirichlet function of the first prior parameter to generate a polynomial probability distribution corresponding to all topics for each word in the training set word set, and obtain the topic word probability distribution;

For each text in the training set, generate a second prior parameter based on the text map of the training set and its corresponding label information, use the Dirichlet function of the second prior parameter to generate a topic distribution, and use the topic distribution as a multinomial distribution parameters, obtain the topic number corresponding to each word in each text, and use the word distribution corresponding to each topic number as a parameter of the multinomial distribution to generate words;

Based on the generated words, the model parameters and the topic word probability distribution are updated.

Further, the generation method of the text-level word representation is:

Based on the text graph of the text to be recognized, a first-layer gated graph neural network is used to merge the representation of each text graph node with the information from its first-order neighbor nodes, and update the embedded representation of each word;

Based on the updated embedding representation of each word, a second layer of gated graph neural network is used to obtain the text-level word representation of each word.

Further, the generation method of the text representation is:

Calculate the attention value of each text word relative to each label based on the text-level word representations and the embedded representations of all labels;

Use the attention value to perform weighted summation on the embedded representations of all tags to obtain the tag semantic component of each word;

Based on the label semantic component of each word, a bidirectional long short-term memory layer is used to obtain the label representation of each word;

The label representation of the word and the text-level word representation are spliced and weighted to obtain weighted features;

Based on the weighted features, max pooling, summing and averaging operations are performed to obtain text representations.

A second aspect of the present invention provides a marketing text recognition system comprising:

A preprocessing module, which is configured to: obtain the text to be recognized, and perform preprocessing;

The graph building module is configured to: construct a text graph of the to-be-recognized text based on the pre-processed to-be-recognized text;

a joint learning module, which is configured to: generate a text-level word representation based on a text graph of the text to be recognized, and combine the embedded representations of all tags to generate a text representation;

A classification module, which is configured to: based on the text representation, use a classifier to obtain a result of whether the text to be recognized belongs to the marketing text;

Among them, the method of obtaining the embedded representation of the label is: based on the text map of the training set and its labels, generate the probability distribution of the topic words, map the probability distribution of the topic words to the label vector space, and learn the correlation between the labels under the guidance of the label map. Sexual relationship and semantic information, get the embedded representation of the label.

A third aspect of the present invention provides a computer-readable storage medium on which a computer program is stored, and when the program is executed by a processor, implements the steps in the above-mentioned method for recognizing marketing text.

A fourth aspect of the present invention provides a computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, the processor implementing the above-mentioned one when executing the program Steps in a Marketing Text Recognition Method.

Compared with the prior art, the beneficial effects of the present invention are:

The invention provides a marketing text recognition method, which provides a topic a priori self-adaptive marked Dirichlet distribution topic model, which directly corresponds each topic to a label, and assigns document topics according to the label set of each document. The space is mapped to a k-dimensional space to modify the Di r i ch l et topic prior, generating topic word probability distributions covering global semantic information.

The present invention provides a marketing text recognition method, which provides a label information integration method. When generating label embedding, the word topic probability distribution is mapped to a real-valued vector, and the statistics of the label are transferred between the real-valued vectors. Co-occurrence information, integrates the global semantics and relevance of tags into the tag representation, and obtains the tag embedding representation that covers the global information.

The invention provides a marketing text recognition method, which acquires word-specific label representations and performs joint learning with the word representations, which can extract more semantic information between labels and texts and improve the model classification effect.

Description of drawings

The accompanying drawings forming a part of the present invention are used to provide further understanding of the present invention, and the exemplary embodiments of the present invention and their descriptions are used to explain the present invention, and do not constitute an improper limitation of the present invention.

1 is a flowchart of a marketing text recognition method according to Embodiment 1 of the present invention;

FIG. 2 is a diagram of a topic model of Dirichlet distribution of topic prior adaptation according to Embodiment 1 of the present invention.

Detailed ways

The present invention will be further described below with reference to the accompanying drawings and embodiments.

It should be noted that the following detailed description is exemplary and intended to provide further explanation of the invention. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.

It should be noted that the terminology used herein is for the purpose of describing specific embodiments only, and is not intended to limit the exemplary embodiments according to the present invention. As used herein, unless the context clearly dictates otherwise, the singular is intended to include the plural as well, furthermore, it is to be understood that when the terms "comprising" and/or "including" are used in this specification, it indicates that There are features, steps, operations, devices, components and/or combinations thereof.

Terminology Explanation:

Label Embedding Representation: Represents the labels of text in vector form for manipulation.

Word Embedding Representation: Word Embedding is actually a class of techniques, a single word is represented as a real vector in a predefined vector space, and each word is mapped to a vector, the word embedding representation.

Example 1

This embodiment provides a marketing text recognition method. First, an independent text graph is constructed for each document, and a label graph is constructed by using the label information of the training set; then a topic model is trained by using the label information corresponding to the text graph nodes and the text graph, Generate the topic word probability distribution; then build a perceptron to map the topic word probability distribution to the label vector space, build a graph convolution network, learn the label correlation and semantic information under the guidance of the label graph, and update the label vector to obtain the label embedding representation; Finally, after jointly learning words and labels to obtain the final text representation, the final classification result is output through the sigmoid layer. As shown in Figure 1, it specifically includes the following steps:

Step 1: Obtain the training set and the text to be recognized, and perform data preprocessing on the training set and the text to be recognized respectively. The preprocessing includes cleaning non-text data, removing stop words, removing low-frequency words, and removing high-frequency words and words. Shape reduction, specifically:

Step (101), clean up non-text data, the data may contain non-text content such as HTML tags, URL addresses, etc., these contents are not helpful for classification or even have a negative impact;

Step (102): Remove stop words. Auxiliary verbs, function words, punctuation marks, etc. are abundant in all articles and cannot reflect the meaning of the text. Usually, removing these stop words from the text can make the model better Fit semantic features to improve the generalization ability of the model;

Step (103), remove low-frequency words and high-frequency words, the importance of the word is proportional to the number of times it appears in the text and inversely proportional to its frequency in the corpus, so the high-frequency words and low-frequency words in the corpus are proportional to Words are not helpful for classification and can be removed (this step is not required in short texts);

Step (104), lemmatization, words such as is, are, and am are completely different contents to the computer, but they have the same meaning, and they are unified through lemmatization to facilitate subsequent processing and analysis.

Step 2: Construct a graph: For the text to be recognized, a text graph is constructed for each text to be recognized based on the preprocessed text to be recognized; for the training set, based on the text in the preprocessed training set, a text map is constructed for each text in the training set. Text builds a text graph, and based on all the labels in the training set, builds a label graph. specific:

Construct a text map, and construct a text map according to the number of co-occurrences of words in a sliding window of a certain size: For a text to be recognized or a certain text in the training set, count the number of co-occurrences of words in a fixed sliding window, and take each unique word as a text map Vertices, the number of co-occurrences between words is the edge of the text graph, and an independent text graph is constructed for each text; the text graph vertices are also called text graph nodes.

Based on the label information (each text in the training set has its corresponding label set, that is, the set of the category to which the text belongs), a label map is constructed, and the label map is constructed according to the number of label co-occurrences in a sliding window of a certain size: For all labels, count the co-occurrence times of labels in the fixed sliding window, and construct a label graph with the category corresponding to each label as the vertex and the label co-occurrence times as the edge.

Among them, the label map and the text map of all texts in the training set are used to train the marketing text recognition model; the marketing text recognition model includes a label embedding representation acquisition module, a word label joint learning module and a classification module. Input the text image of the text to be recognized into the word label joint learning module and the classification module of the trained marketing text recognition model, and obtain the label of the text to be recognized (that is, the result of whether the text to be recognized belongs to the marketing text), that is, the category to which it belongs (excluding the Articles with authentic and credible content, articles with neutral opinions, articles containing advertising content, etc.), specifically, based on the text graph of the text to be recognized, two gated graph neural networks connected in sequence are used to obtain text-level words Representation; based on the text-level word representation, combined with the embedded representation of all labels, the text representation is obtained; based on the text representation, the classifier is used to obtain the text classification result.

Step 3: The label embedding represents the acquisition module, based on the text map of the training set and its corresponding label information, generates the probability distribution of the topic words, maps the probability distribution of the topic words to the label vector space, and learns between labels under the guidance of the label map. The correlation relationship and semantic information of , get the label embedding representation. Including the acquisition of topic word probability distribution module and label information integration module.

(1) Obtain the topic word probability distribution module: First, build a topic-prior adaptive labelled Dirichlet distribution topic model, as shown in Figure 2, and then use Collapsed Gibbs sampling to train topics in the process of generating all documents in the training set model to get the probability distribution of topic words.

The topic-prior adaptive labelled Dirichlet distribution topic model is a probabilistic graphical model. Each label in the model corresponds directly to a topic. For any topic k, a multinomial distribution with parameter η is used to generate the word distribution for it. β _k , for any document d, update α ^(d) through matrix operations so that the prior probability of all topics changes with the change of the document label set, and use the Dirichlet function with parameter α ^(d) to generate topics for it Distribution θ ^(d) ; for the jth word in document d, get its topic number z _j,d ~Multinomial(θ ^(d) ) from the topic distribution θ ^(d ), and then get word w from the topic number _j,d ~Multinomial(βz _j,d ), and repeat the above process until all documents are generated.

The construction and training process of the topic-prior adaptive labelled Dirichlet distribution topic model is as follows:

Step 1: Establish a one-to-one correspondence between labels and topics. The number of label categories in the dataset is equal to the number of topics K; among them, the label refers to the category to which the article belongs, and the topic refers to several words that can summarize the article; The Dirichlet function with the test parameter η generates a multinomial probability distribution corresponding to all topics for each word in the training set word set, and obtains the word distribution corresponding to each topic (topic word probability distribution) β _k = (β _k,1 ,...,β _k,V-1 ,β _k,V )～Dir(.|η), where V also represents the number of words in the entire training set text, β _k,v represents the vth word from the kth word The probability of generating a topic, Dir (that is, Dirchlet) represents the Dirichlet function, η is the first prior parameter, β _k represents the word distribution corresponding to the k-th topic, k∈{1,...,K};

Step 2: For each text in the training set, generate a second prior parameter based on the text map of the training set and its corresponding label information, specifically: for each text d, d∈{1,...,D} , when its label is unknown, use the Bernoulli distribution with label prior Φ to generate the label Λ ^(d) of the text, when the label is known, initialize a label matrix Λ ^(d) with dimension 1*k for it, the text The corresponding position of the owned label is set to 1, otherwise it is set to 0; the initialization dimension is equal to the number of topics. The identity matrix q ^(d) , q ^(d) and the label matrix Λ ^(d) of the text are summed with the second first. Perform multiplication operation on the priori parameter α, that is, the second prior α ^(d) = (q ^(d) + Λ ^(d) )×α; through the above operations, assign a high topic prior to the topic corresponding to the text label set, The remaining topics are assigned a low topic prior to update the second prior parameter α ^(d) ;

Step ③: Use the Dirichlet function with the second prior parameter α ^(d) to generate the topic distribution θ ^(d) , θ ^(d) ~ Dirichlet(.|α ^(d) );

Step 4: Take the topic distribution θ ^(d) as the parameter of the multinomial distribution, and obtain the topic number z _j,d ～Multinomial(θ ^(d) ) corresponding to the jth word in the text d, j∈{1,... , N _d }, N _d is the number of words in the text d;

Step ⑤: The word distribution βz _j,d of the topic corresponding to the topic number z _j,d is used as a parameter of the multinomial distribution to generate the word w _j,d ~Multinomial(βz _j,d );

Step ⑥: Based on all the words w _j,d , use Collapsed Gibbs sampling to train the topic model, update the model parameters (including the first prior parameter η, the second prior parameter α, the topic distribution θ) and the topic word probability distribution, with With the update of parameters, β is also updated, and steps ①-⑤ are repeated continuously, and the probability distribution β of topic words of all documents is obtained after training.

(2) Label information integration module: The label information integration module includes the generation of the label embedded representation module. Generate a label embedding representation module, map the topic word probability distribution to the label vector space through a layer of perceptrons, and then use a two-layer graph convolution network, under the guidance of the label map, learn label correlation and semantic information and integrate them into into the label vector to get the embedded representation of the label.

Step 1: Build a multi-layer perceptron (MLP) activated by relu, map the topic word probability distribution β to the label embedding space, and obtain the embedding representation of all labels covering the global information;

Step 2: Construct the first layer of graph convolutional network (GCN), under the guidance of the label graph, transfer label correlation and semantic information between neighbor nodes, update the embedding representation of the label, and the updated label embedding is represented as H ¹ ;

Step 3: Construct a second-layer graph convolutional network (GCN), transfer high-order information between label nodes, enhance the label embedding representation, and the enhanced label embedding is represented as H ² ;

The learning process of the graph convolutional network is:

H ^l+1 =σ( _AL H ^l W ^l ) (1)

Among them, H ^l is the component representation of the previous layer, H ^l+1 is the output of the current layer, σ is the LeakyReLU activation function, W ^l is the trainable parameter, A _L is the adjacency matrix of label map normalization, and the normalization process is:

为标签图的邻接矩阵表示，

代表邻接矩阵的第i行第j列的值。in,

is the adjacency matrix representation of the label graph,

Represents the value of the ith row and jth column of the adjacency matrix.

Step 4: Word label joint learning module: The word label joint learning module includes obtaining text-level word representation components, extracting word-specific label semantic components, bidirectional long short-term memory layer, soft attention layer and readout layer; obtaining text-level word representations The component uses a two-layer gated graph neural network (GGNN) for text-level information interaction, updates word embedding representation, and then obtains text-level word representation; extracts word-specific label semantic components, and calculates text words corresponding to all labels through attention mechanism. attention value, and use the attention value to weight the label embedding to obtain word-specific label semantic components; bidirectional long short-term memory layer, build a bidirectional long short-term memory network (BiLSTM) to further study the dependencies and words in word-specific label representations Label semantic information to generate word-specific label representations; the soft attention layer weights the concatenated representations of text-level word embedding representations and word-specific label representations to obtain weighted features; readout layer firstly weights the weighted features in a certain dimension The summation is averaged, then the weighted features are max-pooled, and the results of the two-step operations are merged together to obtain a complete text representation.

Step (401): Based on the text graph of the text to be recognized, a first-layer gated graph neural network is used to learn the text graph in units of text, and the text graph nodes merge themselves with features such as semantics from first-order neighbor nodes. information, update the embedding representation of each word;

Step (402): Based on the updated embedded representation of each word, a second-layer gated graph neural network is used to transmit high-level information between each text graph node to obtain a text-level word representation X;

The learning process of the gated graph neural network is:

a ^t =A _T h ^t-1 W _a (4)

z ^t =σ(W _z a ^t +U _z h ^t-1 +b _z ) (5)

r ^t =σ(W _r a ^t +U _r h ^t-1 + _br ) (6)

h ^{~ t} = tanh(W _h a ^t +U _h (r ^t ⊙h ^t-1 )+b _h ) (7)

h ^t =h ^~t ⊙z ^t +h ^t-1 ⊙(1-z ^t ) (8)

Among them, h ^t represents the word embedding representation of the t-th layer gated neural network, which represents the embedding representation of all words in a text graph, W, U, b are trainable parameters, σ is the sigmoid function, and ⊙ represents the dot product Operation, A _T is the adjacency matrix of text graph normalization, the normalization process is similar to the label graph adjacency matrix normalization process, the vector representation h ^t-1 corresponding to the text graph node is the input of the t-1 layer gated graph neural network. When t=2, X=h ² , when t in the formula is 1, it is the first layer of gated graph neural network, and when t is 2, it is the second layer of gated neural network, here is the general expression of gated neural network , only two layers are used in the model described in this paper. X is the text-level word representation learned by the two-layer gated graph neural network. Step (403): Calculate the attention value μ of each text word relative to each label in units of text using equation (9);

表示第j个增强的标签嵌入表示，X_i表示待识别文本中第i个单词的的文本级单词表示，exp()为指数函数，()^T为转置操作。where ^H2 is the enhanced label embedding representation,

represents the j-th enhanced label embedding representation, X _i represents the text-level word representation of the i-th word in the text to be recognized, exp() is an exponential function, and () ^T is a transpose operation.

Step (404): Based on the attention value, use Equation ( ¹⁰ ) to perform a weighted sum of the embedded representations H of all tags to obtain the tag semantic component Q specific to each word;

Step (405): use equation (11) to construct a bidirectional long-term and short-term memory layer, further learn the dependencies and semantic information of word-specific tag semantic components, and generate all word-specific tag representations R;

R=BiLSTM(Q) (11)

where μ _ij is the attention value of the ith word in the text relative to the j th label, and Q _i represents the label semantic component specific to the ith word.

Step (406): splicing all word-specific label representations R to text-level word representations X of all words along the second dimension of the matrix to obtain T;

Step (407): use equation (12) to calculate the soft attention value of T, and use the soft attention value to weight the spliced content to obtain the weighted feature _HV ;

H _V =σ(f _att (T))⊙tanh(f _tr (T)) (12)

Step (408): use equation (13) to perform maximum pooling, summation and averaging operations on the weighted content and combine the results of the operations to obtain H _G ;

Among them, _fatt and _ftr are perceptrons with activation functions of sigmoid and relu, respectively, the former represents soft attention and the latter is a nonlinear transformation, Max represents the maximum pooling operation, V represents a total of V weighted features, and h _v is The vth weighted feature in H _V.

Step 5: The classifier uses the sigmoid function to process the obtained text representation, and obtains the final classification result: H _G obtains the text through the sigmoid function. Results for one or more of the tags, such as articles with ad content.

In this embodiment, the sliding window is an area of a fixed length that is read from an article in order from left to right each time. The fixed length is the size of the sliding window, and the distance of each move is the step size of the sliding window.

In this embodiment, the word co-occurrence information is the number of co-occurrences of words in a sliding window of a certain size.

In this embodiment, the tag co-occurrence information is the number of co-occurrences of tags in a sliding window of a certain size.

In this embodiment, the topic word probability distribution is the word probability distribution corresponding to any topic.

In this embodiment, Collapsed Gibbs samples a Markov Chain Monte Carlo algorithm, typically used for statistical inference.

In this embodiment, the perceptron is a type of neural network.

In this embodiment, the tag correlation relationship refers to a relationship in which tags are mutually dependent or mutually exclusive.

In this embodiment, the graph convolutional network is an application of the convolutional network in machine learning to graph data.

In this embodiment, the gated graph neural network is a GRU-based classical spatial domain information transfer model.

In this embodiment, the high-order information refers to a stacked t-layer gated graph neural network, the information of each node can reach the node t-hops away from it, and this process transmits the high-order information.

In this embodiment, the text-level word representation refers to using a two-layer gated graph neural network to learn a text graph, and updating the word embedding representation after text-level information interaction, and the obtained new representation is the text-level word embedding representation.

In this embodiment, concatenating refers to adding one array to another array along a specific direction axis after a specific dimension.

In this embodiment, the soft attention will be re-weighted in an adaptive manner before all information is aggregated, so that important information can be separated, and the information can be prevented from being interfered by unimportant information.

In this embodiment, the maximum pooling is to take the point with the largest value in the local receptive field.

Embodiment 2

This embodiment provides a marketing text recognition system, which specifically includes the following modules:

A preprocessing module, which is configured to: obtain the text to be recognized, and perform preprocessing;

The graph building module is configured to: construct a text graph of the to-be-recognized text based on the pre-processed to-be-recognized text;

a joint learning module, which is configured to: generate a text-level word representation based on a text graph of the text to be recognized, and combine the embedded representations of all tags to generate a text representation;

A classification module, which is configured to: based on the text representation, use a classifier to obtain a result of whether the text to be recognized belongs to the marketing text;

Among them, the method of obtaining the embedded representation of the label is: based on the text map of the training set and its labels, generate the probability distribution of the topic words, map the probability distribution of the topic words to the label vector space, and learn the correlation between the labels under the guidance of the label map. Sexual relationship and semantic information, get the embedded representation of the label.

It should be noted here that each module in this embodiment corresponds to each step in Embodiment 1 one by one, and the specific implementation process thereof is the same, which is not repeated here.

Embodiment 3

This embodiment provides a computer-readable storage medium on which a computer program is stored, and when the program is executed by a processor, implements the steps in the marketing text recognition method described in the first embodiment above.

Embodiment 4

This embodiment provides a computer device, including a memory, a processor, and a computer program stored in the memory and running on the processor, where the processor implements the one described in the first embodiment when the processor executes the program. Steps in a Marketing Text Recognition Method.

As will be appreciated by one skilled in the art, embodiments of the present invention may be provided as a method, system, or computer program product. Accordingly, the invention may take the form of a hardware embodiment, a software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product embodied on one or more computer-usable storage media having computer-usable program code embodied therein, including but not limited to disk storage, optical storage, and the like.

The present invention is described with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each flow and/or block in the flowchart illustrations and/or block diagrams, and combinations of flows and/or blocks in the flowchart illustrations and/or block diagrams, can be implemented by computer program instructions. These computer program instructions may be provided to the processor of a general purpose computer, special purpose computer, embedded processor or other programmable data processing device to produce a machine such that the instructions executed by the processor of the computer or other programmable data processing device produce Means for implementing the functions specified in a flow or flow of a flowchart and/or a block or blocks of a block diagram.

These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory result in an article of manufacture comprising instruction means, the instructions The apparatus implements the functions specified in the flow or flow of the flowcharts and/or the block or blocks of the block diagrams.

These computer program instructions can also be loaded on a computer or other programmable data processing device to cause a series of operational steps to be performed on the computer or other programmable device to produce a computer-implemented process such that The instructions provide steps for implementing the functions specified in the flow or blocks of the flowcharts and/or the block or blocks of the block diagrams.

Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be implemented by instructing relevant hardware through a computer program, and the program can be stored in a computer-readable storage medium. During execution, the processes of the embodiments of the above-mentioned methods may be included. The storage medium may be a magnetic disk, an optical disk, a read-only memory (Read-Only Memory, ROM), or a random access memory (Random Access Memory, RAM) or the like.

The above descriptions are only preferred embodiments of the present invention, and are not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and changes. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims (10)

1. A marketing text recognition method, comprising:

acquiring a text to be recognized and preprocessing the text;

constructing a text graph of the text to be recognized based on the preprocessed text to be recognized;

generating text level word representation based on a text graph of a text to be recognized, and generating text representation by combining the embedded representations of all labels;

based on the text representation, a classifier is adopted to obtain a result of whether the text to be identified belongs to the marketing text;

the method for acquiring the embedded representation of the label comprises the following steps: generating probability distribution of the subject words based on the text graph and the labels of the training set, mapping the probability distribution of the subject words to a label vector space, and learning the correlation relation and semantic information among the labels under the guidance of the label graph to obtain the embedded representation of the labels.

2. The marketing text recognition method of claim 1, wherein the preprocessing comprises cleaning non-text data, removing stop words, removing low frequency words, removing high frequency words, and morphological restoration.

3. The marketing text recognition method of claim 1, wherein the text graph is constructed by: for a text, counting the co-occurrence times of words in a fixed sliding window, taking each word as the vertex of the text graph, and taking the co-occurrence times between the words as the edges of the text graph.

4. The marketing text recognition method of claim 1, wherein the tag map is constructed by: and for all the labels in the training set, counting the co-occurrence times of the labels in the fixed sliding window, and taking each label as a vertex and the co-occurrence times of the labels as an edge.

5. The marketing text recognition method of claim 1, wherein the topic word probability distribution is generated by:

using a Dirichlet function of a first prior parameter to generate polynomial probability distribution corresponding to all topics for each word in a training set word set to obtain topic word probability distribution;

for each text in the training set, generating a second prior parameter based on a text graph of the training set and corresponding label information of the text graph, generating topic distribution by using a Dirichlet function of the second prior parameter, taking the topic distribution as a parameter of polynomial distribution to obtain a topic number corresponding to each word in each text, and taking word distribution corresponding to each topic number as the parameter of polynomial distribution to generate words;

based on the generated words, the model parameters and the subject word probability distributions are updated.

6. The marketing text recognition method of claim 1, wherein the text-level word representation is generated by:

based on a text graph of a text to be recognized, adopting a first-level gated graph neural network, combining each text graph node and a first-order neighbor node, and updating the embedded representation of each word;

and adopting a second-layer gated graph neural network to obtain a text-level word representation of each word based on the updated embedded representation of each word.

7. The marketing text recognition method of claim 1, wherein the text representation is generated by:

calculating an attention value for each text word relative to each tag based on the text-level word representations and the embedded representations of all tags;

carrying out weighted summation on the embedded representations of all the labels by adopting the attention value to obtain a label semantic component of each word;

obtaining a label representation of each word by adopting a bidirectional long-short term memory layer based on a label semantic component of each word;

the label representation of the word and the text level word representation are spliced and weighted to obtain a weighted feature;

and performing maximum pooling, summation and averaging operation based on the weighting characteristics to obtain text representation.

8. A marketing text recognition system, comprising:

a pre-processing module configured to: acquiring a text to be recognized and preprocessing the text;

a graph building module configured to: constructing a text graph of the text to be recognized based on the preprocessed text to be recognized;

a joint learning module configured to: generating text-level word representation based on a text graph of a text to be recognized, and generating text representation by combining the embedded representations of all tags;

a classification module configured to: based on the text representation, a classifier is adopted to obtain a result of whether the text to be identified belongs to the marketing text;

the method for acquiring the embedded representation of the label comprises the following steps: generating probability distribution of the subject words based on the text graph and the labels of the text graph of the training set, mapping the probability distribution of the subject words to a label vector space, and learning the correlation relation and semantic information among the labels under the guidance of the label graph to obtain the embedded representation of the labels.

9. A computer-readable storage medium, on which a computer program is stored, which, when being executed by a processor, carries out the steps of a marketing text recognition method according to any one of claims 1 to 7.

10. A computer device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor when executing the program implements the steps in a marketing text recognition method of any of claims 1-7.

Images