CN114490935A - Abnormal text detection method, device, computer readable medium and electronic device - Google Patents

Abstract

The application discloses a method and a device for detecting abnormal texts, a computer readable medium and electronic equipment, which can be used in scenes such as artificial intelligence and natural language processing. The method comprises the following steps: acquiring a text to be detected consisting of a plurality of words; extracting features of the text to be detected to obtain a feature sequence of the text to be detected, wherein the feature sequence comprises context features corresponding to a plurality of characters in the text to be detected; mapping the characteristic sequences through a plurality of preset models respectively to obtain processing results corresponding to the preset models; the processing result of the preset model comprises the abnormal probability of the characteristic segment in the characteristic sequence; the lengths of the characteristic segments corresponding to the processing results of different preset models are different; and determining abnormal fragments in the text to be detected according to the abnormal probability of the characteristic fragments indicated by the processing result of each preset model. The text to be detected is detected through the models with various granularities, and accuracy and precision of detection results are improved.

Description

Abnormal text detection method, device, computer readable medium and electronic device

technical field

The present application belongs to the technical field of computers and artificial intelligence, and specifically relates to a method, device, computer-readable medium and electronic device for detecting abnormal text.

Background technique

In many cases, the text data needs to be revised to solve abnormal problems such as typos and semantic and grammatical errors in the text. Currently, in natural language processing, a common detection method is to identify abnormal locations in text through sequence annotation models. Sequence labeling is to give a text sequence, analyze each element in the text sequence, determine the abnormal probability of each element, and finally identify the element with a large abnormal probability as an abnormal element. However, the sequence labeling method usually has the problem of positioning offset, that is, there is a certain gap between the identified abnormal elements and the real abnormal elements. Therefore, the detection accuracy of this detection method is not high and needs to be improved.

It should be noted that the information disclosed in the above Background section is only for enhancing understanding of the background of the application, and therefore may include information that does not form the prior art known to a person of ordinary skill in the art.

SUMMARY OF THE INVENTION

The purpose of the present application is to provide a method, apparatus, computer-readable medium and electronic device for detecting abnormal text, so as to solve the problem of low positioning accuracy of abnormal segments in text in the related art.

Other features and advantages of the present application will become apparent from the following detailed description, or be learned in part by practice of the present application.

According to an aspect of the embodiments of the present application, a method for detecting abnormal text is provided, including:

Get the text to be detected consisting of multiple words;

Perform feature extraction on the text to be detected to obtain a feature sequence of the text to be detected, where the feature sequence includes contextual features corresponding to multiple words in the text to be detected;

The feature sequences are respectively mapped through a plurality of preset models to obtain processing results corresponding to each preset model; wherein, the processing results of the preset models include the abnormal probability of feature segments in the feature sequence, and the The feature segment includes the context feature of at least one word; the lengths of the feature segments corresponding to the processing results of different preset models are different;

According to the abnormal probability of the feature segment indicated by the processing result of each preset model, the abnormal segment in the text to be detected is determined.

According to an aspect of the embodiments of the present application, an apparatus for detecting abnormal text is provided, including:

The text acquisition module is used to acquire the text to be detected composed of multiple words;

a feature extraction module, configured to perform feature extraction on the text to be detected to obtain a feature sequence of the text to be detected, where the feature sequence includes contextual features corresponding to multiple words in the text to be detected;

a mapping processing module, configured to perform mapping processing on the feature sequence through a plurality of preset models, to obtain processing results corresponding to each preset model; wherein the processing results of the preset model include different features in the feature sequence The abnormal probability of the segment, the feature segment includes the context feature of at least one word; the lengths of the feature segments corresponding to the processing results of different preset models are different;

The abnormal segment determination module is configured to determine the abnormal segment in the text to be detected according to the abnormal probability of the feature segment indicated by the processing result of each preset model.

In an embodiment of the present application, the device further includes:

a sample data acquisition module for acquiring sample data consisting of a plurality of words, the words in the sample data having a first label indicating an abnormal state;

The second label generation module is configured to determine a plurality of sample segments in the sample data according to the length of each preset segment based on a plurality of preset segment lengths, and generate the sample according to the first label corresponding to the sample segment. The fragment assigns a second label indicating an abnormal state;

The model training module is configured to use the sample data with the second label corresponding to each preset segment length as a training sample, and train the neural network model through the training samples to obtain preset models corresponding to each preset segment length.

In an embodiment of the present application, the second label generation module is specifically used for:

A window is set with the preset segment length as the window width, and all words contained in the window in the sample data are used as sample segments, wherein the window starts from the sample data according to the set step size. The start position slides to the end position.

In an embodiment of the present application, the first label includes a normal label and an abnormal label; the second label generation module is further configured to:

A second label for the sample segment is generated based on the total amount of abnormal labels within the window and the window width.

In an embodiment of the present application, in the training process of the neural network model, the cross-entropy between the predicted value of the neural network model for the training sample and the second label of the training sample is used as the loss function to update model parameters of the neural network model based on the loss function.

In an embodiment of the present application, the feature extraction module includes:

The word segmentation unit is used to perform word segmentation processing on the text to be detected to obtain a plurality of words arranged in order, and convert each word in the plurality of words arranged in order into a corresponding word according to a preset dictionary word label, to obtain the word sequence of the text to be detected;

A feature extraction unit, configured to perform context feature extraction on the word sequence to obtain the feature sequence of the text to be detected.

In an embodiment of the present application, the feature extraction unit is specifically used for:

Determine the semantic vector and the position vector corresponding to the word label according to the word label in the word sequence;

Generate a vector to be feature extraction according to the word label in the word sequence and the semantic vector and position vector corresponding to the word label;

Context feature extraction is performed on the vector with feature extraction to obtain the feature sequence of the text to be detected.

In an embodiment of the present application, the mapping processing module includes:

a feature segment determining unit, configured to determine the feature segment in the feature sequence by a sliding window method according to the preset segment length corresponding to the preset model;

an abnormal feature representation acquisition unit, configured to acquire the abnormal feature representation of the feature segment through the convolution layer of the preset model;

An abnormal probability acquisition unit, configured to perform mapping processing on the abnormal feature representation through the fully connected layer of the preset model to obtain the abnormal probability of the feature segment.

In an embodiment of the present application, the abnormal feature representation acquisition unit is specifically configured to:

All context features in the feature segment are fused according to the model parameters of the convolutional layer of the preset model to obtain an abnormal feature representation of the feature segment.

In an embodiment of the present application, the model parameters of the convolution layer include a first weight parameter and a first base value parameter; the abnormal feature representation acquisition unit is further configured to:

Weighted summation is performed on all context features in the feature segment by using the first weight parameter to obtain a weight feature;

The weight feature and the first base value parameter are superimposed to obtain an abnormal feature representation of the feature segment.

In an embodiment of the present application, the fully connected layer of the preset model includes a second weight parameter and a second base value parameter; the abnormal probability acquisition unit is specifically used for:

Multiplying the abnormal feature representation by the second weight parameter and then adding the second base value parameter to obtain the feature to be activated;

The feature to be activated is processed through a preset activation function to obtain the abnormal probability of the feature segment.

In an embodiment of the present application, the abnormal segment determination module is specifically used for:

determining the maximum abnormal probability among the abnormal probabilities of the feature segments indicated by the processing results of each preset model;

A plurality of words indicated by the feature segments corresponding to the maximum abnormal probability are used as abnormal segments in the text to be detected.

According to an aspect of the embodiments of the present application, there is provided a computer-readable medium on which a computer program is stored, and when the computer program is executed by a processor, implements the abnormal text detection method in the above technical solution.

According to an aspect of the embodiments of the present application, there is provided an electronic device, the electronic device comprising: a processor; and a memory for storing executable instructions of the processor; wherein the processor is configured to execute the The executable instructions are used to execute the abnormal text detection method in the above technical solution.

According to one aspect of the embodiments of the present application, there is provided a computer program product or computer program, where the computer program product or computer program includes computer instructions, and the computer instructions are stored in a computer-readable storage medium. The processor of the computer device reads the computer instruction from the computer-readable storage medium, and the processor executes the computer instruction, so that the computer device executes the abnormal text detection method in the above technical solution.

In the technical solutions provided in the embodiments of the present application, the feature sequences are processed separately through a plurality of preset models to obtain processing results, wherein the processing results include the abnormal probability of the feature segments, that is, the text to be detected is divided into multiple segments for processing. Anomaly detection, instead of directly detecting the entire sentence of the text to be detected, makes the local features in the text to be detected fully considered in the process of anomaly detection and improves the detection accuracy; at the same time, due to the feature fragments corresponding to the processing results of different preset models The length of the text is different, which is equivalent to the detection of the text to be detected through a variety of granular models, which further improves the accuracy and precision of the detection results.

It is to be understood that both the foregoing general description and the following detailed description are exemplary and explanatory only and are not limiting of the present application.

Description of drawings

The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the application and together with the description serve to explain the principles of the application. Obviously, the drawings in the following description are only some embodiments of the present application, and for those of ordinary skill in the art, other drawings can also be obtained from these drawings without creative effort.

FIG. 1 schematically shows a block diagram of an exemplary system architecture to which the technical solution of the present application is applied.

FIG. 2 schematically shows a flowchart of a method for detecting abnormal text provided by an embodiment of the present application.

FIG. 3 schematically shows a flowchart of feature extraction for text to be detected provided by an embodiment of the present application.

FIG. 4 schematically shows a schematic diagram of determining the length of a preset segment provided by an embodiment of the present application.

FIG. 5 schematically shows a flowchart of a method for constructing a preset model provided by an embodiment of the present application.

FIG. 6 schematically shows a model structure diagram to which the technical solution of the present application is applied.

FIG. 7 schematically shows an application flow chart of the technical solution of the present application in a scenario.

FIG. 8 schematically shows a structural block diagram of an abnormal text detection apparatus provided by an embodiment of the present application.

FIG. 9 schematically shows a structural block diagram of a computer system suitable for implementing the electronic device of the embodiment of the present application.

Detailed ways

Example embodiments will now be described more fully with reference to the accompanying drawings. Example embodiments, however, can be embodied in various forms and should not be construed as limited to the examples set forth herein; rather, these embodiments are provided so that this application will be thorough and complete, and will fully convey the concept of example embodiments to those skilled in the art.

Furthermore, the described features, structures, or characteristics may be combined in any suitable manner in one or more embodiments. In the following description, numerous specific details are provided in order to give a thorough understanding of the embodiments of the present application. However, those skilled in the art will appreciate that the technical solutions of the present application may be practiced without one or more of the specific details, or other methods, components, devices, steps, etc. may be employed. In other instances, well-known methods, devices, implementations, or operations have not been shown or described in detail to avoid obscuring aspects of the present application.

The block diagrams shown in the figures are merely functional entities and do not necessarily necessarily correspond to physically separate entities. That is, these functional entities may be implemented in software, or in one or more hardware modules or integrated circuits, or in different networks and/or processor devices and/or microcontroller devices entity.

The flowcharts shown in the figures are only exemplary illustrations and do not necessarily include all contents and operations/steps, nor do they have to be performed in the order described. For example, some operations/steps can be decomposed, and some operations/steps can be combined or partially combined, so the actual execution order may be changed according to the actual situation.

FIG. 1 schematically shows a block diagram of an exemplary system architecture to which the technical solution of the present application is applied.

As shown in FIG. 1 , the system architecture 100 may include a terminal device 110 , a network 120 and a server 130 . The terminal device 110 may include a smart phone, a tablet computer, a notebook computer, a smart voice interaction device, a smart home appliance, a vehicle-mounted terminal, and the like. The server 130 may be an independent physical server, a server cluster or a distributed system composed of multiple physical servers, or a cloud server providing cloud computing services. The network 120 may be a communication medium of various connection types capable of providing a communication link between the terminal device 110 and the server 130, such as a wired communication link or a wireless communication link.

According to implementation requirements, the system architecture in this embodiment of the present application may have any number of terminal devices, networks, and servers. For example, the server 130 may be a server group composed of a plurality of server devices. In addition, the technical solutions provided in the embodiments of the present application may be applied to the terminal device 110 or the server 130 , or may be jointly implemented by the terminal device 110 and the server 130 , which is not specifically limited in the present application.

The abnormal text detection method provided by the embodiment of the present application is executed by the server 130 , and accordingly, the abnormal text detection apparatus is set in the server 130 . However, those skilled in the art can easily understand that the method for detecting abnormal text provided in the embodiment of the present application may also be executed by the terminal device 110, and correspondingly, the device for detecting abnormal text may also be set in the terminal device 110. This exemplary implementation There is no special restriction on this in the example.

For example, the server 130 acquires text to be detected consisting of multiple characters, and then performs feature extraction on the text to be detected to obtain a feature sequence of the text to be detected, where the feature sequence includes contextual features corresponding to multiple characters in the text to be detected. Next, the server 130 performs mapping processing on the feature sequences through the plurality of preset models, respectively, to obtain processing results corresponding to the respective preset models; wherein, the lengths of the feature segments corresponding to the processing results of different preset models are different. In addition, the processing result of a preset model includes the abnormal probability of multiple feature segments in the feature sequence, and the feature segment includes the context feature of at least one word. Finally, the server 130 determines the abnormal segment in the text to be detected according to the abnormal probability of the feature segment indicated by the processing result of each preset model.

In an embodiment of the present application, after determining the abnormal segment in the text to be detected, the server 130 can return the abnormal segment in the text to be detected to the terminal device 110 through the network 120, and the terminal device 110 can mark the place in the display interface For abnormal segments in the text to be detected, for example, by highlighting the abnormal segments in the text to be detected, the abnormal segments in the text to be detected can be quickly and conveniently known through the display interface of the terminal device 110 .

The technical solutions provided in the embodiments of the present application can be implemented by artificial intelligence technology, such as generating a preset model by artificial intelligence technology. Artificial Intelligence (AI) is a theory, method, technology and application system that uses digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results. In other words, artificial intelligence is a comprehensive technique of computer science that attempts to understand the essence of intelligence and produce a new kind of intelligent machine that can respond in a similar way to human intelligence. Artificial intelligence is to study the design principles and implementation methods of various intelligent machines, so that the machines have the functions of perception, reasoning and decision-making.

Artificial intelligence technology is a comprehensive discipline, involving a wide range of fields, including both hardware-level technology and software-level technology. The basic technologies of artificial intelligence generally include technologies such as sensors, special artificial intelligence chips, cloud computing, distributed storage, big data processing technology, operation/interaction systems, and mechatronics. Artificial intelligence software technology mainly includes computer vision technology, speech processing technology, natural language processing technology, and machine learning/deep learning, autonomous driving, and smart transportation.

The method for detecting abnormal text provided by the present application will be described in detail below with reference to specific embodiments.

FIG. 2 schematically shows a flowchart of a method for detecting abnormal text provided by an embodiment of the present application. The method can be implemented by a terminal device, such as the terminal device 110 shown in FIG. 1; the method can also be implemented by a server, Server 130 as shown in FIG. 1 . As shown in FIG. 2 , the method for detecting abnormal text provided by this embodiment of the present application includes steps 210 to 240, and the details are as follows:

Step 210: Acquire a text to be detected consisting of multiple characters.

Specifically, the text to be detected consists of multiple words, which may be one sentence or multiple sentences. The text to be detected is the text data that has been determined to be abnormal but the abnormal position is not clear. The abnormal situation of the text data includes typos, grammatical errors, and semantic errors in the text data. The text to be detected can be content obtained from text data, for example, a title or body sentence in an article that is confirmed to be abnormal. The text to be detected may also be abnormal text data obtained by performing speech recognition on voice data, or abnormal text data obtained by performing text recognition on image data, which is not limited in this embodiment of the present application. Among them, confirming that the text data is abnormal text can be realized by a trained text detection model.

Step 220: Perform feature extraction on the text to be detected to obtain a feature sequence of the text to be detected, where the feature sequence includes contextual features corresponding to multiple words in the text to be detected.

Specifically, to perform feature extraction on the text to be detected is to extract the contextual semantic features of the text to be detected, convert the text to be detected from text into feature vectors, and obtain contextual features corresponding to multiple words in the text to be detected to form a feature sequence. In the context feature of a word, the semantic information of the word in the text to be detected is included, that is, the abnormal information of the word in the text to be detected is included.

In an embodiment of the present application, as shown in FIG. 3 , the process of performing feature extraction on the text to be detected includes steps 310 to 320, specifically:

Step 310: Perform word segmentation processing on the text to be detected to obtain a plurality of words arranged in order, and convert each word of the plurality of words arranged in order into a corresponding word label according to a preset dictionary, and obtain the text to be detected. word sequence.

Specifically, the word segmentation process is to segment the words in the text to be detected to obtain multiple words arranged in sequence, and the arrangement order of the multiple words is the arrangement order of the words in the text to be detected. Since the text cannot be processed directly, after dividing and obtaining multiple words, each word needs to be converted into a corresponding word label, so as to obtain multiple word labels sorted in order, that is, the word sequence of the text to be detected. . The word tag is a kind of identification of the word, which is equivalent to the ID of the word, which is recorded as Token.

In an embodiment of the present application, words can be converted into corresponding word labels according to a preset dictionary. The default dictionary contains a large number of words and their corresponding word tags. Traverse multiple words arranged in sequence, and for each word, look up the same word as the word in the preset dictionary, and use the word label corresponding to the same word as the word label of the word.

In an embodiment of the present application, when the text to be detected includes multiple sentences, in order to identify the sentences, a sentence start identifier [CLS] can be set at the head of the text to be detected, and a sentence can be set at the end of each sentence tail identifier [SEP]. Generally speaking, the head of the text to be detected is before the start word of the text to be detected. There will be punctuation marks at the end of sentences, so you can first identify the punctuation marks in the sentence, and then set the sentence end identifier at the punctuation mark. In one case, when there is no punctuation mark and no other words after a word, the word can be considered as a sentence end, and a sentence end identifier is set after the word. Of course, when the text to be detected has only one sentence, it is sufficient to set a sentence end identifier after the last word label. Then, the final word sequence is composed of sentence start identifier, word label and sentence end identifier.

Exemplarily, the text to be detected is "I am Chinese, I love China", and the head of the text to be detected is before "I", so the sentence start identifier [CLS] is set before the word label of "I". The "," in the text to be detected is regarded as the end of the sentence, and the "country" in "I love China" is regarded as the end of the sentence, so two sentence end identifiers [SEP] need to be set. After conversion, the word sequence is obtained: [CLS]Token1Token2 Token3 Token4Token5[SEP]Token6 Token7Token8 Token9[SEP], where the token number only represents the order number of the corresponding word in the text to be detected.

Step 320 , perform context feature extraction on the word sequence to obtain a feature sequence of the text to be detected.

Specifically, the context feature is a feature that can represent information such as the context and semantics of the word, and can also reflect the abnormal information of the location of the word.

In an embodiment of the present application, a pre-trained language model may be used to perform context feature extraction on a word sequence, and the pre-trained language model may be a BERT (Bidirectional Encoder Representation from Transformers, bidirectional encoder representation of a transformer) model.

A pretrained language model is a type of model in natural language processing. Natural Language Processing (NLP) is an important direction in the field of computer science and artificial intelligence. It studies various theories and methods that can realize effective communication between humans and computers using natural language. Natural language processing is a science that integrates linguistics, computer science, and mathematics. Therefore, research in this field will involve natural language, the language that people use on a daily basis, so it is closely related to the study of linguistics. Natural language processing technology usually includes text processing, semantic understanding, machine translation, robot question answering, knowledge graph and other technologies.

In an embodiment of the present application, the process of performing context feature extraction on a word sequence specifically includes: determining a semantic vector and a position vector corresponding to the word label according to the word label in the word sequence; The semantic vector and position vector are generated to generate the feature extraction vector; the context feature extraction is performed on the feature extraction vector to obtain the feature sequence of the text to be detected.

Specifically, the semantic vector of the word represents the information after fusion of the global semantic information of the text to be detected and the semantic information of the word. For example, the semantic vector of the word can indicate the sentence in which the word is located (for example, the text to be detected includes sentence A and sentence B, and the word The semantic vector of can indicate whether the word is in sentence A or sentence B), the type of sentence in which the word is located (such as title or body), and the semantic vector is determined by the pre-trained language model according to the word label and the text to be detected. The position vector of the word represents the position information of the word in the text to be detected. This is because the semantic information carried by words in different positions in the text to be detected will be different. Therefore, the position vector of the word is added to make the context feature extraction process. Considering the position information of words, the context feature extraction is more accurate. The position vector is determined by the pretrained language model from the word labels and the text to be detected.

After the semantic vector and position vector corresponding to the word label in the word sequence are determined, the word label is superimposed with the corresponding semantic vector and position vector to obtain a vector to be feature extraction. Exemplarily, for the word sequence: [CLS]Token1Token2 Token3 Token4Token5[SEP], the semantic vector corresponding to each word label (in the order of the word sequence) is: EC E1 E2 E3 E4E5 ES, the position vector corresponding to each word label ( According to the order in the word sequence) is: PC P1P2 P3 P4P5 PS, then the generated vector to be extracted is: [CLS]+EC+PC Token1+E1+P1 Token2+E2+P2Token3+E3+P3 Token4+E4+P4Token5 +E5+P5[SEP]+E6+P6.

Finally, context feature extraction is performed on the feature extraction vector to obtain a feature sequence. The feature sequence includes the context feature of each word, and the context feature of the i-th word is denoted as h _i , then the feature sequence corresponding to the text to be detected with n words is h ₁ h ₂ h ₃ ... h _i ... h _n .

Continuing to refer to FIG. 2, in step 230, the feature sequences are respectively mapped through a plurality of preset models to obtain processing results corresponding to each preset model; wherein, the processing results of the preset models include the abnormal probability of feature segments in the feature sequence, The feature segment includes a context feature of at least one word; the lengths of the feature segments corresponding to the processing results of different preset models are different.

Specifically, the preset model is used to predict the abnormal situation of the feature sequence, and the processing result obtained by the preset model mapping the feature sequence includes the abnormal probability of multiple feature segments. That is, the preset model divides the feature sequence into for multiple feature segments, and then predict the anomaly probability of each feature segment. A feature segment corresponds to a part of a feature sequence, so a feature segment includes context features of at least one word.

In the embodiment of the present application, the feature sequences are respectively mapped through a plurality of preset models. In the processing results obtained by each preset model, the lengths of the feature segments are different, and the length of the feature segments refers to the length of the feature segments. The number of contextual features (equivalent to the number of words corresponding to feature segments). Exemplarily, in this application, five preset models are used to map the feature sequences respectively, the length of the feature segment corresponding to the first preset model is 1, the length of the feature segment corresponding to the second preset model is 2, and the length of the feature segment corresponding to the second preset model is 2. The length of the feature segment corresponding to the three preset models is 3, the length of the feature segment corresponding to the fourth preset model is 4, and the length of the feature segment corresponding to the fifth preset model is 5.

In an embodiment of the present application, the processing process of the feature sequence by the preset model is: according to the preset segment length corresponding to the preset model, determine the feature segment in the feature sequence by the sliding window method, and perform mapping processing on the feature segment , get the abnormal probability of the feature segment.

Specifically, the preset model requires that the lengths of the feature segments into which the feature sequence is divided are preset, that is, the preset segment lengths. Take the preset segment length as the width of a window, and then slide the window along the feature sequence. During the window sliding process, the segment of the feature sequence within the window is the feature segment. Therefore, the feature sequence is divided by the sliding window method. for multiple feature segments. After the feature segments are obtained by division, the feature segments are mapped to obtain the abnormal probability of the feature segments.

During the window sliding process, the window slides with a set step size, and the window header is used as the calculation starting point, then the distance between the current window header and the previous window header is the set step size. Generally, the step size is set to 1, that is to say, the window moves backward by a distance of one word each time, and the window header slides from the start position of the feature sequence to the end position. Then, when dividing the feature sequence, when the preset segment length is k (k>1), set the feature sequence to be h ₁ h ₂ h ₃ ... h _i ... h _n , then the feature segment is [ _hi :h _i+k ], indicating that the feature segment is from the context feature hi of the _ith word to the context feature hi+k of the (i ₊ k)th word, wherein the value of i is from 1 to n. It can be seen that when i=nk, h _i+k is h _n , and when i increases again, i+k will be greater than n, and hi _+k can be replaced by 0 at this time. When the preset segment length is 1, it actually divides the context features in the feature sequence, that is, the text to be detected is divided into single characters, and the obtained feature fragment corresponds to a word in the text to be detected. contextual features. Exemplarily, as shown in Figure 4, taking the preset segment length as 2 as an example, the feature sequence is h ₁ h ₂ h ₃ h ₄ h ₅ , and the set step size of the window sliding is 1, the feature segment can be obtained: [ h ₁ :h ₂ ], [h ₂ :h ₃ ], [h ₃ :h ₄ ], [h ₄ :h ₅ ], [h ₅ :0].

In an embodiment of the present application, the process of mapping the feature segment is as follows: obtaining the abnormal feature representation of the feature segment through the convolution layer of the preset model; performing mapping processing on the abnormal feature representation through the fully connected layer of the preset model , get the abnormal probability of the feature segment.

Specifically, the preset model has a convolution layer and a fully connected layer, the convolution layer is used for extracting abnormal feature representation of the feature segment, and the fully connected layer is used for calculating abnormal probability according to the abnormal feature representation. Specifically, the convolutional layer fuses all contextual features in feature segments through model parameters to obtain abnormal feature representations. The model parameters of the convolution layer include a first weight parameter and a first base value parameter. The first weight parameter and the first base value parameter are the model parameters obtained during the model training process. Different preset models have different convolution layer parameters. different model parameters. When performing fusion processing, firstly, the first weight parameter is used to weight and sum all the context features in the feature segment to obtain the weight feature; then the weight feature and the first base value parameter are superimposed to obtain the abnormal feature representation of the feature segment. . Denote the first weight parameter of the convolutional layer of the preset model corresponding to the preset segment length k as W _1k , the first base value parameter as b _1k , and all the context features in the i-th feature segment as [h _i :h _i+k ], then the abnormal feature representation r _ki of the i-th feature segment is as follows:

r _ki =(W _1k [h _i :h _i+k ]+b _1k )

Among them, k represents the preset segment length, r _ki is the abnormal feature representation of the ith feature segment extracted under the preset segment length k, [hi :hi _+k ] represents the _ith feature segment, W _1k , b _1k is the model parameter of the convolutional layer in the preset model corresponding to the preset segment length k. The length of the preset segment is also equivalent to the width of the convolution kernel of the convolution layer. In this application, the feature sequences are actually processed separately through convolution models of various granularities.

After the abnormal feature representation is obtained, the abnormal feature representation is mapped through the fully connected layer to obtain the abnormal probability of the feature segment. Specifically, the fully connected layer has a second weight parameter and a second base value parameter. First, the abnormal feature representation is multiplied by the second weight parameter, and then added with the second base value parameter to obtain the feature to be activated; The preset activation function set by the fully connected layer processes the feature to be activated to obtain the abnormal probability of the feature segment. The preset activation function can be a ReLU function, a Sigmoid function, a Softmax function, a Linear function, etc., which can be selected according to actual needs during setting. Exemplarily, the second weight parameter of the fully connected layer of the preset model corresponding to the preset segment length k is denoted as W _2k , the second base value parameter is denoted as b _2k , and the Softmax function is used as the preset activation function, then the feature segment The abnormal probability of is as follows:

p _ki =softmax(W _2k r _ki +b _2k )

Among them, p _ki is the abnormal probability of the i-th feature segment extracted under the preset segment length k, r _ki is the abnormal feature representation of the i-th feature segment extracted under the preset segment length k, W _2k , b _2k is the model parameter of the fully connected layer in the preset model corresponding to the preset segment length k.

In an embodiment of the present application, before the feature sequence is mapped by the preset model, a process of constructing the preset model is also included. As shown in FIG. 5 , the process includes steps 510 to 530, specifically:

Step 510: Obtain sample data consisting of a plurality of words, and the words in the sample data have a first label indicating an abnormal state.

Specifically, the sample data is abnormal text data marked with abnormal information, and the abnormal text data is also composed of a plurality of words, and each word has a first label indicating an abnormal state. The abnormal state of a word refers to whether the word is in an abnormal state. The abnormal state can be reflected by different label information of the first label. For example, if the first label is marked as 0, it means that the word is not in an abnormal state (that is, the word is normal.) state), so the first tag of this type can be recorded as a normal tag. The first label is marked as 1, which means that the word is in an abnormal state, so the first label of this type can be recorded as an abnormal label.

In an embodiment of the present application, for the same sample data, different labeling methods may cause the first labels of each word in the sample data to be different. For example, for the sample data "the status of pets is high", the labeler 1 may think that "all people" are not smooth (that is, abnormal), so the two words "all people" are marked as 1, and the rest of the words are marked as 0; Annotator 2 may think that the word "Ren" is redundant, then mark the word "Ren" as 1, and mark the other words as 0; Annotator 3 may think that the word "Du" is not smooth, then mark the word "Du" as 1, and the other words are marked as 1. The word mark is 0. In this way, for the sample data "the status of pets is high", three labeling situations can be obtained as shown in the following table:

Table 1

Step 520: Determine multiple sample segments in the sample data according to each preset segment length based on the multiple preset segment lengths, and assign a second label indicating an abnormal state to the sample segment according to the first label corresponding to the sample segment.

Specifically, a plurality of preset segment lengths are set, the sample data is divided for each preset segment length, a plurality of sample segments corresponding to the sample data are obtained, and the sample segment is determined according to the first label of each word in the sample segment Assign a second label. The second label is calculated according to the first label, and is used to indicate the abnormal condition of the sample segment.

The process of obtaining multiple sample segments of the sample data according to the preset segment length is as follows: setting a window with the preset segment length as the window width, and taking all the words contained in the window in the sample data as the sample segments, wherein the window is based on Set the step size to slide from the start bit to the end bit of the sample data. That is to say, set a window with the preset segment length as the window width, and then slide the window header from the start position of the sample data to the end position according to the set step size. All words within form a sample segment, resulting in multiple sample segments. When the number of words in the window is not enough, it is filled with 0. Generally, set the step size to 1. For the process of obtaining sample fragments, refer to the previous process of obtaining feature fragments, and the two processes are similar. Exemplarily, taking the preset segment length as 2 and the sample data in Table 1 as an example, sample segments can be obtained: pet, object's, location, status, location, capital, people, height, and height 0.

Since the first label of each word in the sample data has only two cases, 0 and 1, when the first label is directly used for model training, there will be a large error between the predicted value of the model and the loss function calculated by the first label. . For example, taking the labeling of person 2 in the above table 1 as an example, the first label of the word "person" is 1, but the model will predict the probability of the word "du" is relatively large, which will lead to a large loss of the model . Then when the model gradient backpropagation updates the parameters for retraining, it will mislead the model to make the probability of the word "du" very small. This leads to model confusion, which reduces the prediction accuracy of the model.

Considering the above situation, the present application assigns a second label to the sample fragment again, and during the window movement process, the second label of the sample fragment is generated according to the total amount of abnormal labels in the window and the width of the window. The second label of the sample segment includes two parts: a normal flag and an abnormal flag, wherein the normal flag is used to represent the probability that the sample segment is in a normal state, the abnormal flag is used to represent the probability that the sample segment is in an abnormal state, the normal flag and the abnormal flag are used. The sum is 1, then, the abnormal flag is determined, and the normal flag is also determined.

In the embodiment of the present application, during the window movement process, the ratio of the total amount of abnormal labels in the window to the window width is used as the abnormality identifier in the second label of the sample segment, and the normal identifier of the sample segment is 1 minus the abnormality logo. When the window width is 1 (ie, the preset segment length is 1), the sample segment is a single word in the sample data, so the second label of the sample segment is the same as the first label of each word. When the window width is 2, taking the labeling of person 2 in the above Table 1 as an example, the sample segment and its corresponding second label (expressed in the form of (abnormal label, normal label)) can be obtained as: pet: (0 ,1), material: (0,1), location: (0,1), status: (0,1), location: (0,1), people: (0.5,0.5), height : (0.5, 0.5), high 0: (0, 1), wherein, pet: (0, 1) means that the abnormal identification of the sample fragment "pet" is 0, and the normal identification is 1. When the window width is 3, taking the labeling of person 2 in Table 1 above as an example, the sample segment and its corresponding second label can be obtained as: pet's: (0,1), object's land: (0,1 ), status: (0,1), status are: (0,1), position are people: (0.33, 0.67), people are tall: (0.33, 0.67), height 0: (0.33, 0.67), High 00: (0,1). It can be seen that the data in the second label no longer has only two values, and the data in the second label is more gentle, which can effectively alleviate the influence of the error in the first label on the model.

Step 530: Use the sample data with the second label corresponding to each preset segment length as a training sample, and train the neural network model through the training sample to obtain a preset model corresponding to each preset segment length.

Specifically, after the second label is assigned to the sample segment in the sample data, the sample data can be used as a training sample to train the neural network model. Through the processing of the foregoing steps, a sample data with a second label can be obtained for each preset segment length, that is, each preset segment length corresponds to a training sample, and in the training process, the preset segment length corresponding to Training the neural network model of the preset segment length with the training samples, that is, to obtain the preset model corresponding to the preset segment length.

In an embodiment of the present application, during the training process of the neural network model, the cross-entropy between the predicted value of the neural network model for the training sample and the second label of the training sample is used as a loss function, and the neural network is updated based on the loss function. Model parameters of the model, the model parameters are parameters such as W _1k , b _1k , W _2k , and b _2k in the preceding steps.

Specifically, the calculation method of the loss function Loss _k is as follows:

Among them, y _0ki represents the normal label in the ith sample segment under the preset segment length k, y _1ki represents the abnormal label in the ith sample segment under the preset segment length k, y _0ki +y _1ki =1; p _0ki represents the probability that the ith sample segment predicted by the neural network model with the preset segment length k is normal, and p _1ki represents the probability that the ith sample segment predicted by the neural network model with the preset segment length k is abnormal.

In an embodiment of the present application, when training the neural network models corresponding to each preset segment length, an iterative training method may be used, that is, each model is sequentially trained in order of the preset segment lengths from small to large.

In an embodiment of the present application, the preset model can also be obtained by training other suitable machine learning models. Machine Learning (ML) is a multi-field interdisciplinary subject involving probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory and other disciplines. It specializes in how computers simulate or realize human learning behaviors to acquire new knowledge or skills, and to reorganize existing knowledge structures to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental way to make computers intelligent, and its applications are in all fields of artificial intelligence. Machine learning and deep learning usually include artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, teaching learning and other technologies.

Continuing to refer to FIG. 2 , step 240 , according to the abnormal probability of the feature segment indicated by the processing result of each preset model, determine the abnormal segment in the text to be detected.

Specifically, the processing result of one preset model includes abnormal probabilities of multiple feature segments of the same length, the processing results of multiple preset models include the abnormal probabilities of multiple feature segments of multiple lengths, and the multiple abnormal probabilities are determined. The maximum abnormal probability in , and the multiple words indicated by the feature segment corresponding to the maximum abnormal probability are used as abnormal segments in the text to be detected, thereby determining the length and position of the abnormal segments in the text to be detected. For example, if the feature segment corresponding to the maximum abnormal probability is the i-th feature segment of the preset segment length k, it is determined that the abnormal segment in the text to be detected is a segment composed of the i-th word to the (i+k)-th word.

In the technical solutions provided in the embodiments of the present application, the feature sequences are processed separately through a plurality of preset models to obtain processing results, wherein the processing results include the abnormal probability of the feature segments, that is, the text to be detected is divided into multiple segments for processing. Anomaly detection, instead of directly detecting the entire sentence of the text to be detected, makes the local features in the text to be detected fully considered in the process of anomaly detection and improves the detection accuracy; at the same time, due to the feature fragments corresponding to the processing results of different preset models The length of the text is different, which is equivalent to the detection of the text to be detected through a variety of granular models, which further improves the accuracy and precision of the detection results.

FIG. 6 schematically shows a model structure diagram to which the technical solution of the present application is applied. As shown in Figure 6, the model structure includes:

Text embedding module (TokenEMBEDDING) 610, this module is used to perform word segmentation processing on the text to be detected, so as to convert the text to be detected 611 into a word sequence composed of word tags (tokens). For details, refer to the relevant description of the aforementioned step 310. This will not be repeated here.

The vector superposition module (TASKEMBEDDING) 620 is used for superimposing the word labels in the word sequence with the corresponding semantic vectors and position vectors to generate a vector to be feature extraction 621 . For details, reference may be made to the relevant description of the foregoing step 320, which will not be repeated here.

The BERT model (BERT MODEL) 630, the BERT model is a pre-trained language model, is used to perform context feature extraction on the feature extraction vector, and output a feature sequence 631.

The multi-granularity convolution module 640 includes 5 kinds of gram convolution models, wherein the gram is the size of the convolution kernel, that is, the preset segment length. In the embodiments of the present application, the granularities of the five convolution models are: 1, 2, 3, 4, and 5, respectively. The convolution model of each granularity performs mapping processing on the output feature sequence 631 respectively to obtain the abnormal probability of the feature segment in the feature sequence, and the length of the feature segment is the same as the granularity of the corresponding convolution model. The abnormal probability of the feature segment is equivalent to the predicted score (score) of the feature segment by the convolution model, and finally the maximum value (MAXscore) is selected from each score to determine the abnormal segment in the text 611 to be detected.

FIG. 7 schematically shows an application flow chart of the technical solution of the present application in a scenario. As shown in Figure 7, the process includes:

S710 , acquiring the incoherent text. The incomprehensible text is the abnormal text, and the abnormal segment in the incomprehensible text can be determined through the technical solution of the present application.

S720. Input the non-fluent text into the non-fluent segment detection model. The non-smooth segment detection model is a model for implementing the technical solution of the present application, that is, the non-smooth segment detection model performs feature extraction on the non-sequential segments to obtain feature sequences; and then maps the feature sequences through a plurality of preset models to obtain each The processing result corresponding to the preset model, wherein the processing result of the preset model includes the abnormal probability of the feature segment in the feature sequence, and the feature segment includes the context feature of at least one word; the length of the feature segment corresponding to the processing results of different preset models different. Finally, the abnormal segment in the text to be detected is determined according to the maximum value of the abnormal probability of the feature segment indicated by the processing result of each preset model.

S730, locating the incompatibility segment according to the model prediction result. The specific positions of the unsmooth segments are determined according to the output result of the unscheduled segment detection model.

S740, a machine audit system. The positioning results are entered into a machine review system, which can be reviewed manually.

S750, the unsmooth segment is highlighted. Enables objects to more quickly identify sloppy segments of sloppy text with highlighting.

It should be noted that although the various steps of the methods of the present application are depicted in the figures in a particular order, this does not require or imply that the steps must be performed in that particular order, or that all illustrated steps must be performed to achieve the desired the result of. Additionally or alternatively, certain steps may be omitted, multiple steps may be combined into one step for execution, and/or one step may be decomposed into multiple steps for execution, and the like.

The following describes the apparatus embodiments of the present application, which can be used to execute the abnormal text detection method in the above-mentioned embodiments of the present application. FIG. 8 schematically shows a structural block diagram of an abnormal text detection apparatus provided by an embodiment of the present application. As shown in FIG. 8 , the apparatus for detecting abnormal text provided by the embodiment of the present application includes:

a text acquisition module 810, configured to acquire text to be detected consisting of multiple characters;

A feature extraction module 820, configured to perform feature extraction on the text to be detected to obtain a feature sequence of the text to be detected, where the feature sequence includes contextual features corresponding to multiple words in the text to be detected;

The mapping processing module 830 is configured to perform mapping processing on the feature sequence through a plurality of preset models, respectively, to obtain processing results corresponding to each preset model; wherein, the processing results of the preset model include different features in the feature sequence. The abnormal probability of the feature segment, the feature segment includes the context feature of at least one word; the lengths of the feature segments corresponding to the processing results of different preset models are different;

The abnormal segment determination module 840 is configured to determine the abnormal segment in the text to be detected according to the abnormal probability of the feature segment indicated by the processing result of each preset model.

In an embodiment of the present application, the device further includes:

a sample data acquisition module for acquiring sample data consisting of a plurality of words, the words in the sample data having a first label indicating an abnormal state;

The second label generation module is configured to determine a plurality of sample segments in the sample data according to the length of each preset segment based on a plurality of preset segment lengths, and generate the sample according to the first label corresponding to the sample segment. The fragment assigns a second label indicating an abnormal state;

The model training module is configured to use the sample data with the second label corresponding to each preset segment length as a training sample, and train the neural network model through the training samples to obtain preset models corresponding to each preset segment length.

In an embodiment of the present application, the second label generation module is specifically used for:

A window is set with the preset segment length as the window width, and all words contained in the window in the sample data are used as sample segments, wherein the window starts from the sample data according to the set step size. The start position slides to the end position.

In an embodiment of the present application, the first label includes a normal label and an abnormal label; the second label generation module is further configured to:

A second label for the sample segment is generated based on the total amount of abnormal labels within the window and the window width.

In an embodiment of the present application, in the training process of the neural network model, the cross-entropy between the predicted value of the neural network model for the training sample and the second label of the training sample is used as the loss function to update model parameters of the neural network model based on the loss function.

In one embodiment of the present application, the feature extraction module 820 includes:

The word segmentation unit is used to perform word segmentation processing on the text to be detected to obtain a plurality of words arranged in order, and convert each word in the plurality of words arranged in order into a corresponding word according to a preset dictionary word label, to obtain the word sequence of the text to be detected;

A feature extraction unit, configured to perform context feature extraction on the word sequence to obtain the feature sequence of the text to be detected.

In an embodiment of the present application, the feature extraction unit is specifically used for:

Determine the semantic vector and the position vector corresponding to the word label according to the word label in the word sequence;

Generate a vector to be feature extraction according to the word label in the word sequence and the semantic vector and position vector corresponding to the word label;

Context feature extraction is performed on the vector with feature extraction to obtain the feature sequence of the text to be detected.

In an embodiment of the present application, the mapping processing module 830 includes:

a feature segment determining unit, configured to determine the feature segment in the feature sequence by a sliding window method according to the preset segment length corresponding to the preset model;

an abnormal feature representation acquisition unit, configured to acquire the abnormal feature representation of the feature segment through the convolution layer of the preset model;

An abnormal probability acquisition unit, configured to perform mapping processing on the abnormal feature representation through the fully connected layer of the preset model to obtain the abnormal probability of the feature segment.

In an embodiment of the present application, the abnormal feature representation acquisition unit is specifically configured to:

All context features in the feature segment are fused according to the model parameters of the convolutional layer of the preset model to obtain an abnormal feature representation of the feature segment.

In an embodiment of the present application, the model parameters of the convolution layer include a first weight parameter and a first base value parameter; the abnormal feature representation acquisition unit is further configured to:

Weighted summation is performed on all context features in the feature segment by using the first weight parameter to obtain a weight feature;

The weight feature and the first base value parameter are superimposed to obtain an abnormal feature representation of the feature segment.

In an embodiment of the present application, the fully connected layer of the preset model includes a second weight parameter and a second base value parameter; the abnormal probability acquisition unit is specifically used for:

Multiplying the abnormal feature representation by the second weight parameter and then adding the second base value parameter to obtain the feature to be activated;

The feature to be activated is processed through a preset activation function to obtain the abnormal probability of the feature segment.

In an embodiment of the present application, the abnormal segment determination module 840 is specifically configured to:

determining the maximum abnormal probability among the abnormal probabilities of the feature segments indicated by the processing results of each preset model;

A plurality of words indicated by the feature segments corresponding to the maximum abnormal probability are used as abnormal segments in the text to be detected.

The specific details of the abnormal text detection apparatus provided in each embodiment of the present application have been described in detail in the corresponding method embodiments, and are not repeated here.

FIG. 9 schematically shows a structural block diagram of a computer system for implementing an electronic device according to an embodiment of the present application.

It should be noted that the computer system 900 of the electronic device shown in FIG. 9 is only an example, and should not impose any limitations on the functions and scope of use of the embodiments of the present application.

As shown in FIG. 9 , the computer system 900 includes a central processing unit 901 (Central Processing Unit, CPU), which can be loaded into a random device according to a program stored in a read-only memory 902 (Read-Only Memory, ROM) or from a storage part 908 Various appropriate operations and processes are executed by accessing a program in a memory 903 (Random Access Memory, RAM). In the random access memory 903, various programs and data necessary for system operation are also stored. The central processing unit 901 , the read-only memory 902 and the random access memory 903 are connected to each other through a bus 904 . An input/output interface 905 (Input/Output interface, ie, I/O interface) is also connected to the bus 904 .

The following components are connected to the input/output interface 905: an input section 906 including a keyboard, a mouse, etc.; an output section 907 including a cathode ray tube (CRT), a liquid crystal display (LCD), etc., and a speaker, etc. ; a storage section 908 including a hard disk, etc.; and a communication section 909 including a network interface card such as a local area network card, a modem, and the like. The communication section 909 performs communication processing via a network such as the Internet. A driver 910 is also connected to the input/output interface 905 as required. A removable medium 911, such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, etc., is mounted on the drive 910 as needed so that a computer program read therefrom is installed into the storage section 908 as needed.

In particular, according to the embodiments of the present application, the processes described in the flowcharts of the respective methods may be implemented as computer software programs. For example, embodiments of the present application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the method illustrated in the flowchart. In such an embodiment, the computer program may be downloaded and installed from the network via the communication portion 909, and/or installed from the removable medium 911. When the computer program is executed by the central processing unit 901, various functions defined in the system of the present application are executed.

It should be noted that the computer-readable medium shown in the embodiments of the present application may be a computer-readable signal medium or a computer-readable storage medium, or any combination of the above two. The computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus or device, or a combination of any of the above. More specific examples of computer readable storage media may include, but are not limited to, electrical connections with one or more wires, portable computer disks, hard disks, random access memory (RAM), read only memory (ROM), erasable Erasable Programmable Read Only Memory (EPROM), flash memory, optical fiber, portable Compact Disc Read-Only Memory (CD-ROM), optical storage device, magnetic storage device, or any suitable of the above The combination. In this application, a computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In this application, however, a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code therein. Such propagated data signals may take a variety of forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the foregoing. A computer-readable signal medium can also be any computer-readable medium other than a computer-readable storage medium that can transmit, propagate, or transport the program for use by or in connection with the instruction execution system, apparatus, or device. Program code embodied on a computer-readable medium may be transmitted using any suitable medium, including but not limited to wireless, wired, etc., or any suitable combination of the foregoing.

The flowchart and block diagrams in the Figures illustrate the architecture, functionality, and operation of possible implementations of systems, methods and computer program products according to various embodiments of the present application. In this regard, each block in the flowchart or block diagrams may represent a module, segment, or portion of code that contains one or more logical functions for implementing the specified functions executable instructions. It should also be noted that, in some alternative implementations, the functions noted in the blocks may occur out of the order noted in the figures. For example, two blocks shown in succession may, in fact, be executed substantially concurrently, or the blocks may sometimes be executed in the reverse order, depending upon the functionality involved. It is also noted that each block of the block diagrams or flowchart illustrations, and combinations of blocks in the block diagrams or flowchart illustrations, can be implemented in special purpose hardware-based systems that perform the specified functions or operations, or can be implemented using A combination of dedicated hardware and computer instructions is implemented.

It should be noted that although several modules or units of the apparatus for action performance are mentioned in the above detailed description, this division is not mandatory. Indeed, according to embodiments of the present application, the features and functions of two or more modules or units described above may be embodied in one module or unit. Conversely, the features and functions of one module or unit described above may be further divided into multiple modules or units to be embodied.

From the description of the above embodiments, those skilled in the art can easily understand that the exemplary embodiments described herein may be implemented by software, or may be implemented by software combined with necessary hardware. Therefore, the technical solutions according to the embodiments of the present application may be embodied in the form of software products, and the software products may be stored in a non-volatile storage medium (which may be CD-ROM, U disk, mobile hard disk, etc.) or on the network , which includes several instructions to cause a computing device (which may be a personal computer, a server, a touch terminal, or a network device, etc.) to execute the method according to the embodiments of the present application.

Other embodiments of the present application will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses or adaptations of this application that follow the general principles of this application and include common knowledge or conventional techniques in the technical field not disclosed in this application .

It is to be understood that the present application is not limited to the precise structures described above and illustrated in the accompanying drawings and that various modifications and changes may be made without departing from the scope thereof. The scope of the application is limited only by the appended claims.

Claims (15)

1. a detection method of abnormal text, is characterized in that, comprises: Get the text to be detected consisting of multiple words; Perform feature extraction on the text to be detected to obtain a feature sequence of the text to be detected, where the feature sequence includes contextual features corresponding to multiple words in the text to be detected; The feature sequences are respectively mapped through a plurality of preset models to obtain processing results corresponding to each preset model; wherein, the processing results of the preset models include the abnormal probability of feature segments in the feature sequence, and the The feature segment includes the context feature of at least one word; the lengths of the feature segments corresponding to the processing results of different preset models are different; According to the abnormal probability of the feature segment indicated by the processing result of each preset model, the abnormal segment in the text to be detected is determined. 2. The method for detecting abnormal text according to claim 1, characterized in that, before each preset model is used to map the feature sequence, and the processing result corresponding to each preset model is obtained, the method Also includes: obtaining sample data consisting of a plurality of words, the words in the sample data having a first label indicating an abnormal state; Based on multiple preset segment lengths, multiple sample segments in the sample data are determined according to each preset segment length, and a second label indicating an abnormal state is assigned to the sample segment according to the first label corresponding to the sample segment. Label; The sample data with the second label corresponding to each preset segment length is used as a training sample, and the neural network model is trained through the training sample to obtain a preset model corresponding to each preset segment length. 3. The method for detecting abnormal text according to claim 2, wherein determining a plurality of sample fragments in the sample data according to each preset fragment length, comprising: A window is set with the preset segment length as the window width, and all words contained in the window in the sample data are used as sample segments, wherein the window starts from the sample data according to the set step size. The start position slides to the end position. 4. The method for detecting abnormal text according to claim 3, wherein the first label includes a normal label and an abnormal label; according to the first label corresponding to the sample fragment, the sample fragment is given an indicating abnormal state the second label, including: A second label for the sample segment is generated based on the total amount of abnormal labels within the window and the window width. 5. The method for detecting abnormal text according to claim 3, wherein in the training process of the neural network model, the prediction value of the neural network model for the training sample and the prediction value of the training sample are compared. The cross-entropy between the second labels is used as a loss function, and the model parameters of the neural network model are updated based on the loss function. 6. The method for detecting abnormal text according to claim 1, wherein the feature extraction is performed on the text to be detected to obtain a feature sequence of the text to be detected, comprising: Perform word segmentation processing on the text to be detected to obtain a plurality of words arranged in order, and convert each word in the plurality of words arranged in order into a corresponding word label according to a preset dictionary, and obtain the word sequence of the text to be detected; Context feature extraction is performed on the word sequence to obtain the feature sequence of the text to be detected. 7. The method for detecting abnormal text according to claim 6, wherein the word sequence is subjected to context feature extraction to obtain the feature sequence of the text to be detected, comprising: Determine the semantic vector and the position vector corresponding to the word label according to the word label in the word sequence; Generate a vector to be feature extraction according to the word label in the word sequence and the semantic vector and position vector corresponding to the word label; Context feature extraction is performed on the vector with feature extraction to obtain the feature sequence of the text to be detected. 8. The method for detecting abnormal text according to claim 1, wherein the feature sequences are respectively subjected to mapping processing through a plurality of preset models to obtain processing results corresponding to the respective preset models, comprising: According to the preset segment length corresponding to the preset model, the feature segment in the feature sequence is determined by a sliding window method; Obtain the abnormal feature representation of the feature segment through the convolution layer of the preset model; The abnormal feature representation is mapped through the fully connected layer of the preset model to obtain the abnormal probability of the feature segment. 9. The method for detecting abnormal text according to claim 8, wherein obtaining the abnormal feature representation of the feature segment through the convolution layer of the preset model, comprising: All context features in the feature segment are fused according to the model parameters of the convolutional layer of the preset model to obtain an abnormal feature representation of the feature segment. 10. The method for detecting abnormal text according to claim 9, wherein the model parameters of the convolution layer include a first weight parameter and a first base value parameter; The model parameters fuse all the contextual features in the feature segment to obtain an abnormal feature representation of the feature segment, including: Weighted summation is performed on all context features in the feature segment by using the first weight parameter to obtain a weight feature; The weight feature and the first base value parameter are superimposed to obtain an abnormal feature representation of the feature segment. 11. The abnormal text detection method according to claim 8, wherein the fully connected layer of the preset model comprises a second weight parameter and a second base value parameter; Perform mapping processing on the abnormal feature representation to obtain the abnormal probability of the feature segment, including: Multiplying the abnormal feature representation by the second weight parameter and then adding the second base value parameter to obtain the feature to be activated; The feature to be activated is processed through a preset activation function to obtain the abnormal probability of the feature segment. 12. The method for detecting abnormal text according to claim 1, wherein determining the abnormal segment in the text to be detected according to the abnormal probability of the feature segment indicated by the processing result of each preset model, comprising: determining the maximum abnormal probability among the abnormal probabilities of the feature segments indicated by the processing results of each preset model; A plurality of words indicated by the feature segments corresponding to the maximum abnormal probability are used as abnormal segments in the text to be detected. 13. A detection device for abnormal text, comprising: The text acquisition module is used to acquire the text to be detected composed of multiple words; a feature extraction module, configured to perform feature extraction on the text to be detected to obtain a feature sequence of the text to be detected, where the feature sequence includes contextual features corresponding to multiple words in the text to be detected; a mapping processing module, configured to perform mapping processing on the feature sequence through a plurality of preset models, to obtain processing results corresponding to each preset model; wherein the processing results of the preset model include different features in the feature sequence The abnormal probability of the segment, the feature segment includes the context feature of at least one word; the lengths of the feature segments corresponding to the processing results of different preset models are different; The abnormal segment determination module is configured to determine the abnormal segment in the text to be detected according to the abnormal probability of the feature segment indicated by the processing result of each preset model. 14. A computer-readable medium on which a computer program is stored, characterized in that, when the computer program is executed by a processor, the method for detecting abnormal text according to any one of claims 1 to 12 is implemented. 15. An electronic device, comprising: processor; and a memory for storing executable instructions for the processor; The execution of the executable instructions by the processor causes the electronic device to execute the abnormal text detection method according to any one of claims 1 to 12.

Images