A method and device for identifying public opinion risks

Abstract

The embodiment of the present application provides a method and device for identifying public opinion risks, which can be applied to the field of information extraction technology in natural language processing to improve the accuracy of risk subject category identification. The method includes: obtaining a text to be identified; inputting the text to be identified into a pre-trained risk subject identification model to obtain multiple risk subjects; extracting the contextual semantic information, contextual semantic information and the semantic information of each risk subject itself centered on the position information of each risk subject, and performing convolution operations and global maximum pooling processing on the contextual semantic information, contextual semantic information and the semantic information of each risk subject respectively; determining the risk subject category corresponding to each risk subject based on the results of the convolution operations and global maximum pooling processing.

Description

Technical Field

[0001] The present invention relates to the technical field of information extraction in natural language processing, and in particular to a method and device for identifying public opinion risks. Background Art

[0002] Public opinion refers to the social attitudes of the public as the subject towards the social managers, enterprises, individuals and other organizations as the object and their political, social and moral orientations around the occurrence, development and changes of social events in a certain social space. With the rapid development of Internet technology, the openness and flexibility of the Internet have made it one of the main carriers for reflecting social public opinion.

[0003] Among them, the information extraction task in natural language processing capabilities can quickly and accurately identify the financial text content involving risk public opinion in massive financial news texts, and extract the event subjects corresponding to the risk public opinion. In the bank's risk control department, risk management team and its branch departments, the subjects of the events corresponding to the risk public opinion are extracted and introduced into the risk public opinion early warning modeling many times, helping staff to more effectively carry out risk prediction, risk level assessment, risk avoidance and other tasks.

[0004] At present, the category to which an event subject belongs is mainly identified by concatenating the vector of the text to be identified and the vector of the event subject itself, and identifying the category to which the event subject belongs based on the concatenated vector. However, when there are multiple event subjects in the text to be identified, since the text vectors to be identified corresponding to the multiple event subjects are the same, the accuracy of identifying the category of the event subject based on the concatenated vector is low. Summary of the invention

[0005] The embodiments of the present application provide a method and device for identifying public opinion risks, which are used to improve the accuracy of identifying risk subject categories.

[0006] In a first aspect, a method for identifying public opinion risks is provided, the method comprising:

[0007] Get the text to be recognized;

[0008] Input the text to be identified into a pre-trained risk subject identification model to obtain multiple risk subjects; wherein the risk subject identification model is used to obtain a semantic feature vector of the text to be identified, and based on the semantic feature vector, determine whether a field consisting of each starting position and a corresponding end position is an entity, and determine whether the entity is an entity corresponding to the risk subject;

[0009] Extracting contextual semantic information, contextual semantic information and semantic information of each risk subject centered on the risk subject based on the position information of each risk subject, and performing convolution operation and global maximum pooling processing on the contextual semantic information, contextual semantic information and semantic information of each risk subject respectively;

[0010] The risk subject category corresponding to each risk subject is determined based on the convolution operation and the global maximum pooling processing results.

[0011] Optionally, the obtaining of the semantic feature vector of the text to be recognized includes:

[0012] Obtaining a sentence semantic feature vector and a paragraph semantic feature vector of the text to be recognized;

[0013] The sentence semantic feature vector and the paragraph semantic feature vector are fused based on a preset method to obtain the semantic feature vector.

[0014] Optionally, the fusing the sentence semantic feature vector and the paragraph semantic feature vector based on a preset manner includes:

[0015] Adding the sentence semantic feature vector and the paragraph semantic feature vector; or,

[0016] The sentence semantic feature vector and the paragraph semantic feature vector are fused based on a gating mechanism.

[0017] Optionally, extracting contextual semantic information centered on each risk subject, contextual semantic information, and semantic information of each risk subject itself based on the location information of each risk subject includes:

[0018] Based on a preset boundary correction method, the boundaries of the multiple risk subjects are corrected to obtain multiple target risk subjects;

[0019] Based on the position information of each target risk subject, contextual semantic information centered on each target risk subject, contextual semantic information, and semantic information of each target risk subject itself are extracted.

[0020] Optionally, the risk subject identification model is further used to calculate the score of each field as an entity, and the boundary correction of the multiple risk subjects based on the preset boundary correction method is performed to obtain multiple target risk subjects, including:

[0021] Determine whether there is a boundary conflict between at least two risk subjects among the multiple risk subjects; wherein the boundary conflict indicates that there is an intersection between the fields corresponding to the at least two risk subjects;

[0022] If so, the risk subject with the highest score among the at least two risk subjects is determined as the target risk subject, and the other risk subjects are determined as non-target risk subjects.

[0023] Optionally, the performing boundary correction on the multiple risk subjects based on a preset boundary correction method to obtain multiple target risk subjects includes:

[0024] Determine whether there are two risk subjects in the plurality of risk subjects whose positions are continuous; wherein the continuous position indicates that the end position of the first risk subject of the two risk subjects is continuous with the start position of the second risk subject;

[0025] If so, a risk subject obtained by merging the two risk subjects is determined as the target risk subject.

[0026] Optionally, the method further includes:

[0027] Each target risk subject is matched with the pre-built professional terminology dictionary for the longest common subsequence;

[0028] If the first target risk subject is a subsequence of the first professional term, determining whether the first professional term exists in the text to be identified; wherein the first professional term is a word in the professional term dictionary library;

[0029] If so, the first target risk subject will be replaced by the first professional term.

[0030] Optionally, the loss function used in the risk subject identification model is a focal loss function.

[0031] In a second aspect, a public opinion risk identification device is provided, the device comprising:

[0032] An acquisition module is used to acquire the text to be recognized;

[0033] A processing module, used for inputting the text to be identified into a pre-trained risk subject identification model to obtain multiple risk subjects; wherein the risk subject identification model is used to obtain a semantic feature vector of the text to be identified, and based on the semantic feature vector, determine whether a field consisting of each starting position and a corresponding end position is an entity, and determine whether the entity is an entity corresponding to the risk subject;

[0034] The processing module is further used to extract contextual semantic information, contextual semantic information and semantic information of each risk subject centered on the risk subject based on the position information of each risk subject, and perform convolution operation and global maximum pooling processing on the contextual semantic information, contextual semantic information and semantic information of each risk subject respectively;

[0035] The processing module is also used to determine the risk subject category corresponding to each risk subject based on the convolution operation and the global maximum pooling processing results.

[0036] Optionally, the risk subject identification model is specifically used for:

[0037] Obtaining a sentence semantic feature vector and a paragraph semantic feature vector of the text to be recognized;

[0038] The sentence semantic feature vector and the paragraph semantic feature vector are fused based on a preset method to obtain the semantic feature vector.

[0039] Optionally, the risk subject identification model is specifically used for:

[0040] Adding the sentence semantic feature vector and the paragraph semantic feature vector; or,

[0041] The sentence semantic feature vector and the paragraph semantic feature vector are fused based on a gating mechanism.

[0042] Optionally, the processing module is specifically used to:

[0043] Based on a preset boundary correction method, the boundaries of the multiple risk subjects are corrected to obtain multiple target risk subjects;

[0044] Based on the position information of each target risk subject, contextual semantic information centered on each target risk subject, contextual semantic information, and semantic information of each target risk subject itself are extracted.

[0045] Optionally, the risk subject identification model is further used to calculate the score of each field as an entity, and the processing module is specifically used to:

[0046] Determine whether there is a boundary conflict between at least two risk subjects among the multiple risk subjects; wherein the boundary conflict indicates that there is an intersection between the fields corresponding to the at least two risk subjects;

[0047] If so, the risk subject with the highest score among the at least two risk subjects is determined as the target risk subject, and the other risk subjects are determined as non-target risk subjects.

[0048] Optionally, the processing module is specifically used to:

[0049] Determine whether there are two risk subjects in the plurality of risk subjects whose positions are continuous; wherein the continuous position indicates that the end position of the first risk subject of the two risk subjects is continuous with the start position of the second risk subject;

[0050] If so, a risk subject obtained by merging the two risk subjects is determined as the target risk subject.

[0051] Optionally, the processing module is further used to:

[0052] Each target risk subject is matched with the pre-built professional terminology dictionary for the longest common subsequence;

[0053] If the first target risk subject is a subsequence of the first professional term, determining whether the first professional term exists in the text to be identified; wherein the first professional term is a word in the professional term dictionary library;

[0054] If so, the first target risk subject will be replaced by the first professional term.

[0055] According to a third aspect, an electronic device is provided, including:

[0056] A memory for storing program instructions;

[0057] The processor is used to call the program instructions stored in the memory and execute the steps included in the method described in the first aspect according to the obtained program instructions.

[0058] According to a fourth aspect, a computer-readable storage medium is provided, wherein the computer-readable storage medium is used to store instructions, and when the instructions are executed, the method described in the first aspect is implemented.

[0059] According to a fifth aspect, a computer program product comprising instructions is provided. The computer program product stores instructions which, when executed on a computer, cause the computer to execute the method described in the first aspect.

[0060] In an embodiment of the present application, a text to be identified is obtained, and the text to be identified is input into a risk subject identification model. A semantic feature vector of the text to be identified is obtained through the risk subject identification model. Based on the semantic feature vector, it is determined whether each field consisting of a starting position and a corresponding ending position is an entity, and whether the entity is an entity corresponding to the risk subject. A plurality of risk subjects are obtained, and then, based on the position information of each risk subject, contextual semantic information, contextual semantic information, and the semantic information of each risk subject itself are extracted. Convolution operations and global maximum pooling processing are performed on the contextual semantic information, contextual semantic information, and the semantic information of each risk subject itself obtained. Based on the results of the convolution operations and the global maximum pooling processing, the risk event category corresponding to each risk subject is determined.

[0061] That is to say, the present application obtains the contextual semantic information (or contextual semantic vector), the self-semantic information and the contextual semantic information of each risk subject based on the location information of each risk subject, and performs convolution operation and global maximum pooling operation on the obtained semantic information. Compared with the simple concatenation of two vectors in the prior art, the accuracy of risk subject category identification can be effectively improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for describing the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application.

[0063] Figure 1 A flowchart of a method for identifying public opinion risks provided in an embodiment of the present application;

[0064] Figure 2 A multi-granularity semantic feature network structure diagram provided in an embodiment of the present application;

[0065] Figure 3 A network structure diagram of a Biaffine network provided in an embodiment of the present application;

[0066] Figure 4 A structural diagram of a risk subject classification model provided in an embodiment of the present application;

[0067] Figure 5 A structural block diagram of a public opinion risk identification device provided in an embodiment of the present application;

[0068] Figure 6 A schematic diagram of the structure of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0069] In order to make the purpose, technical scheme and advantages of the present application clearer, the technical scheme in the embodiment of the present application will be clearly and completely described below in conjunction with the drawings in the embodiment of the present application. Obviously, the described embodiment is only a part of the embodiment of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of the present application. In the absence of conflict, the embodiments in the present application and the features in the embodiments can be combined with each other arbitrarily. In addition, although the logical order is shown in the flow chart, in some cases, the steps shown or described can be performed in an order different from that here.

[0070] The terms "first" and "second" in the specification and claims of the present application and the above-mentioned drawings are used to distinguish different objects, rather than to describe a specific order. In addition, the term "comprising" and any of their variations are intended to cover non-exclusive protection. For example, a process, method, system, product or device comprising a series of steps or units is not limited to the listed steps or units, but optionally also includes steps or units that are not listed, or optionally also includes other steps or units inherent to these processes, methods, products or devices. "Multiple" in the present application can mean at least two, for example, two, three or more, and the embodiments of the present application are not limited.

[0071] In addition, the term "and / or" in this article is only a description of the association relationship of associated objects, indicating that there can be three relationships. For example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone. In addition, the character " / " in this article, unless otherwise specified, generally indicates that the associated objects before and after are in an "or" relationship.

[0072] Before introducing the embodiments of the present application, some technical features in the present application are first introduced to facilitate understanding by those skilled in the art.

[0073] Event: historical events, real events or information that have occurred, such as financial events, legal events, social events, historical events, natural events, etc.

[0074] Event subject: refers to a list of names related to the target event, including names of people, places, organizations, institutions, etc., and the names in the list of names are arranged in descending order of importance relative to the target event. Relevant consulting information can be obtained through the event subject to identify risks, assist in decision-making, and determine the development trend of the event, etc., so as to achieve reverse indexing of related information through the event subject. The present invention is oriented to the financial field, and the event subject mentioned mainly refers to the financial event subject, which refers to the organizational entity here.

[0075] Fully connected layer: Each node in the fully connected layer is connected to all nodes in the previous layer to integrate the features extracted previously. The fully connected layer can integrate the discriminative local information in the previous neural network layer. If a softmax layer or sigmoid layer is connected to the fully connected layer for output, the probability distribution of multiple classifications can be obtained, thereby achieving classification.

[0076] Entity recognition layer: Use a certain method to identify names of people, places, organizations, dates, etc. from a text. The entity recognition layer of the present invention is to identify risk subjects.

[0077] Convolutional layer: used to extract features, that is, to perform an inner product operation on the pixels of a block according to multiple certain weights (i.e., convolution kernels), and its output is one of the extracted features.

[0078] Pooling layer: that is, sampling or aggregating a piece of data, for example, selecting the maximum value (or average value) of the area to replace the area.

[0079] The following is a detailed description of the public opinion risk identification method provided by the embodiment of the present application in conjunction with the accompanying drawings. Figure 1 As shown, the flowchart of the public opinion risk identification method provided by this application is described as follows:

[0080] Step 101: Obtain the text to be recognized;

[0081] Among them, since the embodiment of the present application is oriented to the financial field, the text to be identified in the embodiment of the present application can be a financial news text, or other information and messages related to finance.

[0082] Step 102: input the text to be identified into a pre-trained risk subject identification model to obtain multiple risk subjects;

[0083] In an embodiment of the present application, the risk subject identification model includes a multi-granularity semantic feature network, a dual affine classification network and an entity recognition layer, wherein the multi-granularity semantic feature network is used to obtain the semantic feature vector of the text to be identified, the dual affine classification network is used to determine whether the field consisting of each starting position and the corresponding ending position is an entity based on the semantic feature vector of the text to be identified obtained by the multi-granularity semantic feature network, and the entity recognition layer is used to determine whether the entity determined by the dual affine classification network is the entity corresponding to the risk subject.

[0084] As a possible implementation, in an embodiment of the present application, a multi-granularity semantic feature network obtains sentence semantic feature vectors and paragraph semantic feature vectors of the text to be recognized, fuses the sentence semantic feature vectors and the paragraph semantic feature vectors, and obtains a final semantic feature vector.

[0085] Among them, the multi-granularity semantic feature network is built based on the pre-trained language representation model (BidirectionalEncoder Representation from Transformers, BERT) and the long short-term memory network (Long Short-Term Memory, LSTM). The multi-granularity semantic feature network mainly includes three modules, sentence semantic extraction module, paragraph semantic extraction module, sentence and paragraph semantic fusion module. After obtaining the text to be recognized, the text to be recognized is first represented as a vector using word embedding, and then the sentence semantic extraction module uses the sentence-level BILSTM method to encode and extract the sentence semantic feature vector, and the paragraph semantic extraction module uses the paragraph-level BILSTM method to encode and extract the paragraph semantic feature. Finally, the sentence semantic feature vector and the paragraph semantic feature vector are fused through the sentence and paragraph semantic fusion module. The specific network structure is as follows: Figure 2 shown.

[0086] After introducing the specific structure of the multi-granularity semantic feature network of the present application, the principles of extracting sentence semantic feature vectors by the sentence semantic extraction module, extracting paragraph semantic feature vectors by the paragraph semantic extraction module, and fusing sentence semantic feature vectors and paragraph semantic feature vectors are respectively introduced below.

[0087] The principle of extracting sentence semantic feature vectors by the sentence semantic extraction module is as follows:

[0088] Assume that the text to be recognized includes a total of k sentences, and the word sequence of sentence 1 is ..., the sentence k-word sequence is The corresponding word vector sequences of k sentences obtained by word embedding are as follows: Then sentence-level BILSTM (BILSTM sent ) performs semantic encoding on each sentence. The specific formula is as follows:

[0089]

[0090] …

[0091]

[0092] Get a paragraph consisting of k sentences through sentence-level BILSTM sent The semantic representation of

[0093] The principle of extracting paragraph semantic feature vectors by the paragraph semantic extraction module is as follows:

[0094] Continuing with the above example, the text to be recognized includes a total of k sentences, and the sentence 1 word sequence is ..., the sentence k-word sequence is The corresponding word vector sequences of k sentences obtained by word embedding are as follows: Then, paragraph-level BILSTM (i.e., BILSTM para ) performs semantic encoding on each sentence. The specific formula is as follows:

[0095]

[0096] Get a paragraph consisting of k sentences through paragraph-level BILSTM para The semantic representation of

[0097] The principle of fusing sentence semantic feature vectors and paragraph semantic feature vectors is as follows:

[0098] Continuing with the above example, the text to be recognized includes a total of k sentences, and the sentence 1 word sequence is ..., the sentence k-word sequence is The sentence semantic feature vector obtained by the sentence semantic extraction module after k sentences is: The paragraph semantic feature vector obtained by the paragraph semantic extraction module for k sentences is: The sentence semantic feature vector and the paragraph semantic feature vector can be fused in the following two ways.

[0099] The first method is simple summation, which is to add the sentence semantic feature vector and the paragraph semantic feature vector. For example, the i-th word s in the k-th sentence i k The semantic feature vector extracted by the sentence semantic feature extraction module is The paragraph semantic feature vector extracted by the paragraph semantic feature extraction module is Then the semantic feature vector obtained by simply summing the i-th word in the k-th sentence is

[0100] The second method is to fuse the sentence semantic feature vector and the paragraph semantic feature vector based on the gating mechanism. Taking the above as an example, the semantic feature vector of the i-th word in the k-th sentence after fusion based on the gating mechanism is:

[0101]

[0102]

[0103] Among them, * represents dot product, W1, W2 and b are the network parameters learned by the model (sigmoid).

[0104] In the specific implementation process, considering that the prior art directly truncates the text to be recognized, which easily loses the semantic information of the text to be recognized, or directly cuts the text to be recognized into multiple fragments and then summarizes them, which is relatively rough to process, the present application constructs a multi-granularity semantic feature network to fully capture the chapter-level semantic information of the text to be recognized, so that the obtained semantic feature vector of the text to be recognized is more accurate.

[0105] After obtaining the semantic feature vector of the text to be recognized through the multi-granularity semantic feature network, the biaffine classification network (Biaffine network) is used to determine whether the field (span) composed of each start position and the corresponding end position is an entity based on the semantic feature vector. The network structure of the Biaffine network is as follows: Figure 3 As shown in the figure, it can be seen that the Biaffine network uses two independent linear networks, FFNN_start (i.e. FFNN s ) network and FFNN_end (i.e. FFNN e ) network, FFNN_start network and FFNN_end network respectively obtain the starting position vector and the ending position vector of span based on the semantic feature vector obtained by the multi-granularity semantic feature network (which can be represented as vectors h s and h e ), and then perform semantic interaction (Biaffine Classify) based on the start position vector and the end position vector of the span. Then, under the condition that the start position is less than the end position, determine the score of an entity formed by the span of the start position and the end position, and determine whether the span is an entity based on the score.

[0106] For example, suppose the text to be recognized contains a total of k sentences, whose word sequences are s1,…,s n , the word sequence is passed through the multi-granularity semantic feature network to obtain the semantic feature vector x1,…,x n . Apply Biaffine to the semantic feature vector and get the score matrix tensor r m (i.e., each span is the score of an entity), and its dimension is n×n×c, where c is the number of entity categories + 1, r m The calculation formula is as follows:

[0107]

[0108]

[0109] rm (i) = h s (i) T U m h e (i)+W m (h s (i)+h e (i))+b m

[0110] Among them, s i and e i are the starting and ending positions of span i, respectively. and are the semantic feature vector of the starting position of span i and the semantic feature vector of the ending position of span i obtained by the multi-granularity semantic feature network, U m is a d×c×d dimensional tensor, a double affine matrix, and is the network parameter learned by the model, W m is a 2d×c matrix, which is also the network parameter learned by the model, b m It is a 2D vector and is also the network parameter learned by the model.

[0111] After obtaining the score of each span as an entity, it is determined whether the score of each span as an entity is greater than a preset threshold. If the score is greater than the preset threshold, the span is determined to be an entity.

[0112] Furthermore, after determining whether each span is an entity, an entity recognition layer (i.e., a softmax layer) can be spliced ​​behind the dual affine classification network to perform multi-classification to determine whether the entity is the entity corresponding to the risk subject. The specific formula is as follows:

[0113]

[0114] Among them, p m (i c ) is the score of span i on entity type c, and its value is between [0,1].

[0115] In the specific implementation process, considering that the prior art predicts the starting position and ending position of the entity body separately, the starting position token and the ending position token information are not used for interaction, which may lead to the situation that the number of the starting position and the ending position of the entity body are inconsistent, and post-processing is required according to the prediction probability, which makes the entity recognition efficiency low. In the embodiment of the present application, the Biaffine network adopts a bilinear affine mechanism, which allows the entity starting position token and the entity ending position token information to interact, and then combines the entity starting position token and the ending position token information to determine whether the span composed of the starting position token and the ending position token is an entity, which can effectively improve the entity recognition efficiency.

[0116] As a possible implementation method, in the embodiment of the present application, before the text to be identified is input into the pre-trained risk subject identification model, it is also necessary to train the defined network structure through the annotated pre-training corpus, and continuously optimize the loss function through gradient descent to learn various features in the corpus information. Among them, the loss function is used to estimate the degree of inconsistency between the model's predicted value f(x) and the true value Y. In the embodiment of the present application, the risk subject identification model uses the focal loss loss function instead of the general cross entropy loss function. The specific formula of the focal loss loss function is as follows:

[0117]

[0118] Among them, α is used to adjust the importance of positive and negative samples, γ is used to adjust the speed of reducing the weight of simple samples, α and γ are both hyperparameters, and y′ is the model prediction score, which is between [0,1]. During the loss function optimization process, if the prediction score is 0.95, it indicates that the sample is a simple sample. At this time (1-0.95) γ The smaller the value of is, the smaller the corresponding loss value is. If the prediction score is 0.3, it means that the sample is difficult to classify, and its loss value is relatively large. Therefore, the focal loss function can not only solve the problem of imbalance between positive and negative samples, but also reduce the loss of easy-to-classify samples, increase the weight of difficult and misclassified samples, make the model pay more attention to such difficult-to-distinguish samples, and increase the robustness of the model.

[0119] Step 103: extracting contextual semantic information centered on each risk subject, contextual semantic information, and semantic information of each risk subject itself based on the location information of each risk subject;

[0120] As described in step 102, the risk subject identification model has obtained the starting position and ending position of each risk subject in the process of identifying the risk subject. In an embodiment of the present application, the preceding semantic information, following semantic information and the semantic information of the risk subject itself can be respectively obtained based on the position information of each risk subject among the multiple risk subjects identified by the risk subject identification model (i.e., the starting position and ending position of each risk subject).

[0121] Specifically, after the risk subject identification model identifies the text to be identified and obtains multiple risk subjects, the position information corresponding to each risk subject of the multiple risk subjects can be input into the dynamic convolutional neural network to extract the contextual semantic information, contextual semantic information and the semantic information of each risk subject. For example, three dynamic convolutional neural networks are used to extract the contextual semantic information, the semantic information of the current risk subject (span) based on the position information corresponding to each risk subject.

[0122] As mentioned above, the present invention is oriented to the financial field, and the event subjects mentioned mainly refer to financial event subjects, that is, organizational entities, and they are all long entities, and long entities are prone to boundary errors, and entities identified using a dual affine classification network are easily contained in another entity, or one entity is located inside another entity. Therefore, in response to the above situation, the embodiment of the present application can also perform boundary correction on multiple risk subjects based on a preset boundary correction method before extracting the contextual semantic information, contextual semantic information and its own semantic information centered on each risk subject based on the position information of each risk subject, so as to obtain multiple target risk subjects, and then extract the contextual semantic information, contextual semantic information and its own semantic information centered on each target risk subject based on the position information of each target risk subject.

[0123] The following describes a specific implementation method for performing boundary correction on multiple risk entities in an embodiment of the present application.

[0124] A first possible implementation method is to determine whether there are boundary conflicts between at least two risk subjects among multiple risk subjects, that is, to determine whether there are intersections in the fields corresponding to at least two risk subjects among multiple risk subjects (corresponding to the aforementioned one entity being contained in another entity, or one entity being located inside another entity). If there are intersections in the fields corresponding to at least two risk subjects, the scores of each span entity calculated by the risk subject identification model are used to sort the at least two risk subjects in descending order of scores, and the risk subject with the highest score is determined as the target risk subject, and the other risk subjects are determined as non-target risk subjects.

[0125] For example, the starting position of the span1 field in the text to be identified is 13, and the ending position is 18 (i.e., span1 is the 13th-18th characters in the text to be identified), and the score of the field as an entity calculated by the risk subject identification model is 0.95; the starting position of the span2 field in the text to be identified is 15, and the ending position is 18, and the score of the field as an entity calculated by the risk subject identification model is 0.92; the starting position of the span3 field in the text to be identified is 17, and the ending position is 20, and the score of the field as an entity calculated by the risk subject identification model is 0.93. Based on the position information of each field, it can be determined that the three fields have an intersection (i.e., there is a boundary conflict), then the three fields are sorted in descending order according to the scores, and span1>span3>span2 is obtained. At this time, it can be determined that the risk subject corresponding to the span field is the target risk subject, and the risk subjects corresponding to the span2 and span3 fields are non-target risk subjects.

[0126] A second possible implementation method is to determine whether there are two risk subjects among multiple risk subjects whose positions are continuous, that is, to determine whether there are two risk subjects among multiple risk subjects whose end position of the first risk subject is continuous with the starting position of the second risk subject. If so, it indicates that one risk subject may be identified as two risk subjects. At this time, the two risk subjects can be merged, and the risk subject obtained after the merger can be determined as the target risk subject.

[0127] As a possible implementation method, after boundary correction is performed on multiple risk subjects to obtain multiple target risk subjects, each target risk subject can be matched with a pre-built professional terminology dictionary for the longest common subsequence. If the first target risk subject is a subsequence of the first professional term, it is determined whether the first professional term exists in the text to be identified, wherein the first professional term is a word in the professional terminology dictionary. If so, the first target risk subject is replaced with the first professional term.

[0128] For example, the text to be identified is "A Co., Ltd. has declared bankruptcy", and the risk subject identification model identifies the risk subject as A (A is the first target risk subject). When the risk subject identification result "A" is matched with the pre-built professional terminology dictionary for the longest common subsequence, A is determined to be the longest common subsequence of the first professional term. The first professional term can be, for example, "A Co., Ltd." or "A Technology Co., Ltd." At this time, in order to avoid erroneous replacement, it is necessary to further determine whether the first professional term exists in the text to be identified. If the first professional term is "A Co., Ltd.", it is determined that the first professional term exists in the text to be identified, and the first professional term can be used to replace the first target risk subject; if the first professional term is "A Technology Co., Ltd.", it is determined that the first professional term does not exist in the text to be identified, and the first professional term cannot be used to replace the first target risk subject.

[0129] In the specific implementation process, correcting the boundaries of multiple risk subjects identified by the risk subject identification model and replacing the identified target risk subjects with words in the professional terminology dictionary can effectively improve the accuracy of risk subject identification.

[0130] Step 104: performing convolution operation and global maximum pooling processing on the contextual semantic information, contextual semantic information and semantic information of each risk subject respectively;

[0131] After obtaining the contextual semantic information, contextual semantic information and the semantic information of the risk subject corresponding to each risk subject through step 103, the contextual semantic information, contextual semantic information and the semantic information of the risk subject corresponding to each risk subject can be convolved and processed by global maximum pooling. For example, the contextual semantic information, contextual semantic information and the semantic information of the risk subject can be convolved and processed by global maximum pooling (i.e., CNN+MaxPlooing) through a convolution layer and a pooling layer.

[0132] Step 105: Determine the risk subject category corresponding to each risk subject based on the convolution operation and the global maximum pooling processing results.

[0133] Among them, the risk subject categories mainly include: "subject of production suspension and reduction event", "subject of bankruptcy event", "subject of abnormal asset event", "subject of breach of contract and dishonesty event", "subject of reduction of holdings event", "subject of abnormal director, supervisor and member event", "subject of rating deterioration event", "subject of financial fraud event", "subject of asset seizure event", "subject of loss event", etc. In the embodiment of the present application, considering that a risk subject may belong to multiple risk subject categories (for example, simultaneously belonging to the subject of breach of contract and dishonesty event and the subject of production suspension and reduction event), when determining the risk subject category corresponding to each risk subject based on the convolution operation and the global maximum pooling processing results, the classification problem of each risk subject can be converted into a binary classification problem. For example, whether each risk subject belongs to the current risk subject category can be converted into a binary classification problem through a multi-layer perceptron (MultiLayer Perceptron, MLP) and a sigmoid layer (i.e., MLP+sigmoid). Among them, if the number of risk subject categories is k, the number of binary classifiers is also k. Each risk subject is judged whether it belongs to the current risk subject category by k binary classifiers. The specific classification formula is as follows:

[0134]

[0135] Among them, e represents the score of risk subject e, and its value is [0,1].

[0136] In some other embodiments, before extracting the semantic information of each risk subject through the dynamic convolutional neural network and determining the risk subject category corresponding to each risk subject through the MLP and sigmoid layers, the dynamic convolutional neural network and the sigmoid layer can also be constructed to form a risk subject classification model. The structure of the risk subject classification model is as follows: Figure 4 As shown, the risk subject classification model is then trained through the annotated pre-training corpus, and the loss function is continuously optimized through gradient descent to learn various features in the corpus information. Among them, the loss function used by the risk subject classification model is the same as the loss function used by the risk subject identification model, which will not be repeated here.

[0137] In the specific implementation process, this application can fully capture the chapter-level semantic information of the text to be identified by constructing a multi-granularity semantic feature network, so that the obtained semantic feature vector of the text to be identified is more accurate, and by using the focal loss loss function as the loss function in the model training process, it can not only solve the problem of imbalance between positive and negative samples, but also reduce the loss of easy-to-classify samples and increase the weight of difficult and misclassified samples, so that the model pays more attention to such difficult-to-distinguish samples, thereby increasing the robustness of the model, and by performing boundary correction on multiple risk subjects identified by the risk subject identification model, the accuracy of risk subject identification is effectively improved.

[0138] In order to better understand the technical solution of the present application, the pre-distortion extension model and the method for implementing pre-distortion provided by the present application will be explained in conjunction with specific embodiments below.

[0139] Example

[0140] The obtained text to be identified is "Since 2020, among the investment projects of the first phase of the Vision Fund, Brandless and OneWeb have declared bankruptcy, which directly resulted in the loss of all the US$2.24 billion invested by the first phase of the Vision Fund". After inputting it into the risk subject identification model for risk subject identification, two risk subjects are obtained, namely "Brandless" and "OneWeb". Among them, the starting position of "Brandless" in the text to be identified is 22, and the ending position is 30. The starting position of "OneWeb" in the text to be identified is 32, and the ending position is 37. Based on the position information (22, 30) of "Brandless", the semantic information of the field with the starting position 1 and the ending position 21 (the above semantic information) and the starting position 22 and the ending position 31 are extracted respectively. The semantic information of the field (i.e., Brandless) and the semantic information of the field with a starting position of 31 and an ending position of the last character of the text to be recognized (position 75) (the semantic information below), and the semantic information of the field with a starting position of 1 and an ending position of 31 (the semantic information above), the semantic information of the field with a starting position of 32 and an ending position of 37 (i.e., OneWeb), and the semantic information of the field with a starting position of 38 and an ending position of 75 (the semantic information below) are respectively extracted based on the position information of "OneWeb". Convolution operations and global maximum pooling processing are performed on the semantic information extracted based on the position information of Brandless and OneWeb, respectively, and then based on the operation results, the risk entity category corresponding to Brandless and OneWeb is determined to be "bankruptcy event subject".

[0141] Based on the same inventive concept, the embodiment of the present application provides a public opinion risk identification device, which can realize the functions corresponding to the aforementioned public opinion risk identification method. The public opinion risk identification device can be a hardware structure, a software module, or a hardware structure plus a software module. The public opinion risk identification device can be implemented by a chip system, which can be composed of a chip or include a chip and other discrete devices. Please refer to Figure 5 As shown, the public opinion risk identification device includes an acquisition module 501, a processing module 502 and a risk subject identification model 503. Among them:

[0142] An acquisition module 501 is used to acquire text to be recognized;

[0143] Processing module 502, used for inputting the text to be identified into a pre-trained risk subject identification model to obtain multiple risk subjects; wherein the risk subject identification model is used to obtain a semantic feature vector of the text to be identified, and based on the semantic feature vector, determine whether a field consisting of each starting position and a corresponding end position is an entity, and determine whether the entity is an entity corresponding to the risk subject;

[0144] The processing module 502 is further used to extract contextual semantic information, contextual semantic information and semantic information of each risk subject centered on the risk subject based on the position information of each risk subject, and perform convolution operation and global maximum pooling processing on the contextual semantic information, contextual semantic information and semantic information of each risk subject respectively;

[0145] The processing module 502 is further used to determine the risk subject category corresponding to each risk subject based on the convolution operation and the global maximum pooling processing results.

[0146] Optionally, the risk subject identification model 503 is specifically used for:

[0147] Obtaining a sentence semantic feature vector and a paragraph semantic feature vector of the text to be recognized;

[0148] The sentence semantic feature vector and the paragraph semantic feature vector are fused based on a preset method to obtain the semantic feature vector.

[0149] Optionally, the risk subject identification model 503 is specifically used for:

[0150] Adding the sentence semantic feature vector and the paragraph semantic feature vector; or,

[0151] The sentence semantic feature vector and the paragraph semantic feature vector are fused based on a gating mechanism.

[0152] Optionally, the processing module 502 is specifically configured to:

[0153] Based on a preset boundary correction method, the boundaries of the multiple risk subjects are corrected to obtain multiple target risk subjects;

[0154] Based on the position information of each target risk subject, contextual semantic information centered on each target risk subject, contextual semantic information, and semantic information of each target risk subject itself are extracted.

[0155] Optionally, the risk subject identification model 503 is further used to calculate the score of each field as an entity, and the processing module 502 is specifically used to:

[0156] Determine whether there is a boundary conflict between at least two risk subjects among the multiple risk subjects; wherein the boundary conflict indicates that there is an intersection between the fields corresponding to the at least two risk subjects;

[0157] If so, the risk subject with the highest score among the at least two risk subjects is determined as the target risk subject, and the other risk subjects are determined as non-target risk subjects.

[0158] Optionally, the processing module 502 is specifically configured to:

[0159] Determine whether there are two risk subjects in the plurality of risk subjects whose positions are continuous; wherein the continuous position indicates that the end position of the first risk subject of the two risk subjects is continuous with the start position of the second risk subject;

[0160] If so, a risk subject obtained by merging the two risk subjects is determined as the target risk subject.

[0161] Optionally, the processing module 502 is further configured to:

[0162] Each target risk subject is matched with the pre-built professional terminology dictionary for the longest common subsequence;

[0163] If the first target risk subject is a subsequence of the first professional term, determining whether the first professional term exists in the text to be identified; wherein the first professional term is a word in the professional term dictionary library;

[0164] If so, the first target risk subject will be replaced by the first professional term.

[0165] All relevant contents of each step involved in the embodiment of the aforementioned public opinion risk identification method can be referred to the functional description of the functional module corresponding to the public opinion risk identification device in the embodiment of the present application, and will not be repeated here.

[0166] The division of modules in the embodiments of the present application is schematic and is only a logical function division. There may be other division methods in actual implementation. In addition, each functional module in each embodiment of the present application may be integrated into a processor, or may exist physically separately, or two or more modules may be integrated into one module. The above-mentioned integrated modules may be implemented in the form of hardware or in the form of software functional modules.

[0167] Based on the same inventive concept, the present application embodiment provides an electronic device. Figure 6As shown, the electronic device includes at least one processor 601 and a memory 602 connected to the at least one processor. The specific connection medium between the processor 601 and the memory 602 is not limited in the embodiment of the present application. Figure 6 In the example, the processor 601 and the memory 602 are connected via a bus 600. Figure 6 The connection between other components is shown by bold lines, and is not intended to be limiting. The bus 600 can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 6 Only one thick line is used in the diagram, but this does not mean that there is only one bus or only one type of bus.

[0168] In an embodiment of the present application, the memory 602 stores instructions that can be executed by at least one processor 601. By executing the instructions stored in the memory 602, the at least one processor 601 can execute the steps included in the aforementioned public opinion risk identification method.

[0169] Among them, the processor 601 is the control center of the electronic device, which can use various interfaces and lines to connect various parts of the entire electronic device, and monitor the electronic device as a whole by running or executing instructions stored in the memory 602 and calling data stored in the memory 602, various functions of the electronic device and processing data. Optionally, the processor 601 may include one or more processing units, and the processor 601 may integrate an application processor and a modem processor, wherein the application processor mainly processes operating systems and application programs, etc., and the modem processor mainly processes wireless communications. It is understandable that the above-mentioned modem processor may not be integrated into the processor 601. In some embodiments, the processor 601 and the memory 602 may be implemented on the same chip, and in some embodiments, they may also be implemented separately on independent chips.

[0170] Processor 601 can be a general-purpose processor, such as a central processing unit (CPU), a digital signal processor, an application-specific integrated circuit, a field programmable gate array or other programmable logic device, a discrete gate or transistor logic device, a discrete hardware component, and can implement or execute the methods, steps and logic block diagrams disclosed in the embodiments of this application. A general-purpose processor can be a microprocessor or any conventional processor, etc. The steps of the public opinion risk identification method disclosed in the embodiments of this application can be directly reflected as being executed by a hardware processor, or can be executed by a combination of hardware and software modules in the processor.

[0171] The memory 602 is a non-volatile computer-readable storage medium that can be used to store non-volatile software programs, non-volatile computer executable programs and modules. The memory 602 may include at least one type of storage medium, such as a flash memory, a hard disk, a multimedia card, a card-type memory, a random access memory (Random Access Memory, RAM), a static random access memory (Static Random Access Memory, SRAM), a programmable read-only memory (Programmable Read Only Memory, PROM), a read-only memory (Read Only Memory, ROM), an electrically erasable programmable read-only memory (Electrically Erasable Programmable Read-Only Memory, EEPROM), a magnetic memory, a disk, an optical disk, etc. The memory 602 is any other medium that can be used to carry or store a desired program code in the form of an instruction or data structure and can be accessed by a computer, but is not limited thereto. The memory 602 in the embodiment of the present application can also be a circuit or any other device that can realize a storage function, for storing program instructions and / or data.

[0172] By designing and programming the processor 601, the code corresponding to the public opinion risk identification method introduced in the above embodiment can be solidified into the chip, so that the chip can execute the steps of the above public opinion risk identification method during operation. How to design and program the processor 601 is a technology well known to those skilled in the art and will not be repeated here.

[0173] Based on the same inventive concept, an embodiment of the present application also provides a computer-readable storage medium, which stores computer instructions. When the computer instructions are executed on a computer, the computer executes the steps of the public opinion risk identification method as described above.

[0174] In some possible implementations, various aspects of the public opinion risk identification method provided in the present application can also be implemented in the form of a program product, which includes a program code. When the program product is run on an electronic device, the program code is used to enable the detection device to execute the steps of the public opinion risk identification method according to various exemplary embodiments of the present application described above in this specification.

[0175] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code.

[0176] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0177] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0178] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process in the computer or other programmable device. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0179] Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalents, the present application is also intended to include these modifications and variations.

Claims

1. A method for identifying public opinion risks, characterized in that: The method comprises: Get the text to be recognized; The text to be identified is input into a pre-trained risk subject identification model to obtain multiple risk subjects; wherein the risk subject identification model is used to obtain a semantic feature vector of the text to be identified, and based on the semantic feature vector, determine whether a field consisting of each starting position and a corresponding ending position is an entity, and determine whether the entity is an entity corresponding to the risk subject, and the semantic feature vector is obtained by fusing the sentence semantic feature vector and the paragraph semantic feature vector of the text to be identified based on a preset method; Extracting contextual semantic information, contextual semantic information and semantic information of each risk subject centered on the risk subject based on the position information of each risk subject, and performing convolution operation and global maximum pooling processing on the contextual semantic information, contextual semantic information and semantic information of each risk subject respectively; The risk subject category corresponding to each risk subject is determined based on the convolution operation and the global maximum pooling processing results.

2. The method according to claim 1, characterized in that The fusing the sentence semantic feature vector and the paragraph semantic feature vector based on a preset method includes: Adding the sentence semantic feature vector and the paragraph semantic feature vector; or, The sentence semantic feature vector and the paragraph semantic feature vector are fused based on a gating mechanism.

3. The method according to claim 1, characterized in that The extracting, based on the location information of each risk subject, context semantic information centered on each risk subject, and semantic information of each risk subject itself includes: Based on a preset boundary correction method, the boundaries of the multiple risk subjects are corrected to obtain multiple target risk subjects; Based on the position information of each target risk subject, contextual semantic information centered on each target risk subject, contextual semantic information, and semantic information of each target risk subject itself are extracted.

4. The method according to claim 3, characterized in that The risk subject identification model is also used to calculate the score of each field as an entity, and the boundary correction of the multiple risk subjects based on the preset boundary correction method is performed to obtain multiple target risk subjects, including: Determine whether there is a boundary conflict between at least two risk subjects among the multiple risk subjects; wherein the boundary conflict indicates that there is an intersection between the fields corresponding to the at least two risk subjects; If so, the risk subject with the highest score among the at least two risk subjects is determined as the target risk subject, and the other risk subjects are determined as non-target risk subjects.

5. The method according to claim 3, characterized in that The step of performing boundary correction on the plurality of risk subjects based on a preset boundary correction method to obtain a plurality of target risk subjects includes: Determine whether there are two risk subjects in the plurality of risk subjects whose positions are continuous; wherein the continuous position indicates that the end position of the first risk subject of the two risk subjects is continuous with the start position of the second risk subject; If so, a risk subject obtained by merging the two risk subjects is determined as the target risk subject.

6. The method according to any one of claims 3 to 5, characterized in that: The method further comprises: Each target risk subject is matched with the pre-built professional terminology dictionary for the longest common subsequence; If the first target risk subject is a subsequence of the first professional term, determining whether the first professional term exists in the text to be identified; wherein the first professional term is a word in the professional term dictionary library; If so, the first target risk subject will be replaced by the first professional term.

7. The method according to claim 1, characterized in that The loss function used in the risk subject identification model is the focal loss function.

8. A public opinion risk identification device, characterized in that: The device comprises: An acquisition module is used to acquire the text to be recognized; A processing module, used for inputting the text to be identified into a pre-trained risk subject identification model to obtain multiple risk subjects; wherein the risk subject identification model is used to obtain a semantic feature vector of the text to be identified, and based on the semantic feature vector, determine whether a field consisting of each starting position and a corresponding end position is an entity, and determine whether the entity is an entity corresponding to the risk subject, wherein the semantic feature vector is obtained by fusing the sentence semantic feature vector and the paragraph semantic feature vector of the text to be identified based on a preset method; The processing module is further used to extract contextual semantic information, contextual semantic information and semantic information of each risk subject centered on the risk subject based on the position information of each risk subject, and perform convolution operation and global maximum pooling processing on the contextual semantic information, contextual semantic information and semantic information of each risk subject respectively; The processing module is also used to determine the risk subject category corresponding to each risk subject based on the convolution operation and the global maximum pooling processing results.

9. The device according to claim 8, characterized in that The risk subject identification model is specifically used for: Adding the sentence semantic feature vector and the paragraph semantic feature vector; or, The sentence semantic feature vector and the paragraph semantic feature vector are fused based on a gating mechanism.

10. The device according to claim 8, characterized in that The processing module is specifically used for: Based on a preset boundary correction method, the boundaries of the multiple risk subjects are corrected to obtain multiple target risk subjects; Based on the position information of each target risk subject, contextual semantic information centered on each target risk subject, contextual semantic information, and semantic information of each target risk subject itself are extracted.

11. The device according to claim 10, characterized in that The risk subject identification model is also used to calculate the score of each field as an entity. The processing module is specifically used to: Determine whether there is a boundary conflict between at least two risk subjects among the multiple risk subjects; wherein the boundary conflict indicates that there is an intersection between the fields corresponding to the at least two risk subjects; If so, the risk subject with the highest score among the at least two risk subjects is determined as the target risk subject, and the other risk subjects are determined as non-target risk subjects.

12. The device according to claim 10, characterized in that The processing module is specifically used for: Determine whether there are two risk subjects in the plurality of risk subjects whose positions are continuous; wherein the continuous position indicates that the end position of the first risk subject of the two risk subjects is continuous with the start position of the second risk subject; If so, a risk subject obtained by merging the two risk subjects is determined as the target risk subject.

13. The device according to any one of claims 10 to 12, characterized in that: The processing module is further used for: Each target risk subject is matched with the pre-built professional terminology dictionary for the longest common subsequence; If the first target risk subject is a subsequence of the first professional term, determining whether the first professional term exists in the text to be identified; wherein the first professional term is a word in the professional term dictionary library; If so, the first target risk subject will be replaced by the first professional term.

14. An electronic device, characterized in that: include: A memory for storing program instructions; A processor is used to call the program instructions stored in the memory, and execute the steps included in the method according to any one of claims 1 to 7 according to the obtained program instructions.

15. A computer-readable storage medium, characterized in that: The computer-readable storage medium is used to store instructions, and when the instructions are executed, the method according to any one of claims 1 to 7 is implemented.

16. A computer program product comprising instructions, characterized in that The computer program product stores instructions, and when the computer program product is run on a computer, the computer is caused to execute the method according to any one of claims 1 to 7.

Images