CN112507715A

CN112507715A - Method, device, equipment and storage medium for determining incidence relation between entities

Info

Publication number: CN112507715A
Application number: CN202011377531.4A
Authority: CN
Inventors: 张峥; 徐伟建; 罗雨
Original assignee: Beijing Baidu Netcom Science and Technology Co Ltd
Current assignee: Beijing Baidu Netcom Science and Technology Co Ltd
Priority date: 2020-11-30
Filing date: 2020-11-30
Publication date: 2021-03-16
Anticipated expiration: 2040-11-30
Also published as: CN112507715B

Abstract

The application discloses a method, a device, equipment and a storage medium for determining an incidence relation between entities, which are applied to the technical fields of natural language processing, knowledge maps and deep learning. The specific implementation scheme is as follows: acquiring target associated information, wherein the target associated information comprises a first word representing a first entity, a second word representing a second entity and a third word representing a target association relation; determining a first semantic feature according to the target association relation and similar segments aiming at the target association information in a preset segment library; determining a description text aiming at the target associated information according to the target associated information and a preset knowledge graph; determining a second semantic feature according to the target associated information and the description text; and determining a confidence level of the target association relationship between the first entity and the second entity according to the first semantic feature and the second semantic feature.

Description

Method, device, equipment and storage medium for determining incidence relation between entities

Technical Field

The present application relates to the technical field of artificial intelligence, and in particular, to the technical field of natural language processing, knowledge mapping, and deep learning, and more particularly, to a method, an apparatus, a device, and a storage medium for determining an association relationship between entities.

Background

With the development of artificial intelligence technology, electronic management of information is gradually emerging. In each field, in order to improve information acquisition efficiency, information having an association relationship may be associated and managed when the information is electronically managed. Wherein, the association relationship between the information can be determined by the association relationship between the entities described by the information.

In the related art, when information association management is performed, two entities having an association relationship are often recalled from text data in various fields by using a natural language processing method, and the method has a high requirement on the quality of the text data. Moreover, for a specific field, the construction cost of some specific association relations is high, the information density of the specific association relations in the text is low, and the recall rate of recalling the entities with the specific association relations from the text by adopting a natural language processing method is limited.

Disclosure of Invention

A method, device, equipment and storage medium for determining the incidence relation between entities are provided, which are used for improving the accuracy of determining the incidence relation and are beneficial to reducing the construction cost of the target incidence relation.

According to a first aspect, there is provided a method of determining an association between entities, comprising: acquiring target associated information, wherein the target associated information comprises a first word representing a first entity, a second word representing a second entity and a third word representing a target association relation; determining a first semantic feature according to the target associated information and similar segments aiming at the target associated information in a preset segment library; determining a description text aiming at the target associated information according to the target associated information and a preset knowledge graph; determining a second semantic feature according to the target associated information and the description text; and determining a confidence level of the target association relationship between the first entity and the second entity according to the first semantic feature and the second semantic feature.

According to a second aspect, there is provided an apparatus for determining an association relationship between entities, comprising: the information acquisition module is used for acquiring target associated information, wherein the target associated information comprises a first word representing a first entity, a second word representing a second entity and a third word representing a target association relation; the first characteristic determining module is used for determining a first semantic characteristic according to the target associated information and the similar text segments aiming at the target associated information in the preset text segment library; the text determination module is used for determining a description text aiming at the target associated information according to the target associated information and a preset knowledge graph; the second characteristic determining module is used for determining a second semantic characteristic according to the target associated information and the description text; and the confidence coefficient determining module is used for determining the confidence coefficient of the target association relationship between the first entity and the second entity according to the first semantic feature and the second semantic feature.

According to a third aspect, there is provided an electronic device comprising: at least one processor; and a memory communicatively coupled to the at least one processor; the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the method for determining associations between entities provided herein.

According to a fourth aspect, there is provided a non-transitory computer readable storage medium having stored thereon computer instructions for causing a computer to perform the method of determining associations between entities as provided herein.

It should be understood that the statements in this section do not necessarily identify key or critical features of the embodiments of the present disclosure, nor do they limit the scope of the present disclosure. Other features of the present disclosure will become apparent from the following description.

Drawings

The drawings are included to provide a better understanding of the present solution and are not intended to limit the present application. Wherein:

fig. 1 is a schematic application scenario diagram of a method, an apparatus, a device, and a storage medium for determining an association relationship between entities according to an embodiment of the present application;

FIG. 2 is a flowchart illustrating a method for determining associations between entities according to an embodiment of the present disclosure;

FIG. 3 is a schematic diagram of the principle of determining a first semantic feature according to an embodiment of the present application;

FIG. 4 is a schematic diagram of determining a first semantic feature according to another embodiment of the present application;

FIG. 5 is a schematic diagram illustrating a principle of determining a description text for target associated information according to an embodiment of the present application;

FIG. 6 is a schematic diagram illustrating a principle of determining an association relationship between entities according to an embodiment of the present application;

FIG. 7 is a block diagram of an apparatus for determining associations between entities according to an embodiment of the present disclosure; and

FIG. 8 is a block diagram of an electronic device that is used to implement a method for determining associations between entities according to an embodiment of the present application.

Detailed Description

The following description of the exemplary embodiments of the present application, taken in conjunction with the accompanying drawings, includes various details of the embodiments of the application for the understanding of the same, which are to be considered exemplary only. Accordingly, those of ordinary skill in the art will recognize that various changes and modifications of the embodiments described herein can be made without departing from the scope and spirit of the present application. Also, descriptions of well-known functions and constructions are omitted in the following description for clarity and conciseness.

The application provides a method for determining an incidence relation between entities. The method can first obtain target associated information, where the target associated information includes a first word representing a first entity, a second word representing a second entity, and a third word representing a target association relationship. Then, determining a first semantic feature according to the target associated information and similar segments aiming at the target associated information in a preset segment library; and determining a description text aiming at the target associated information according to the target associated information and a preset knowledge graph, and determining a second semantic feature according to the target associated information and the description text. And finally, determining the confidence degree of the target association relationship between the first entity and the second entity according to the first semantic feature and the second semantic feature.

An application scenario of the method and apparatus provided by the present application will be described below with reference to fig. 1.

Fig. 1 is an application scenario diagram of a method, an apparatus, a device, and a storage medium for determining an association relationship between entities according to an embodiment of the present application.

As shown in fig. 1, the application scenario 100 of this embodiment may include, for example, a terminal device 110 and a user 120.

The terminal device 110 may be, for example, various electronic devices capable of providing an interactive interface and having a processing function, including but not limited to, a smart phone, a tablet computer, a laptop portable computer, a desktop computer, and the like.

According to an embodiment of the present application, the terminal device 110 may process the two entity words according to a pre-trained processing model, for example, to obtain an association relationship between two entities represented by the two entity words. Illustratively, the terminal device 110 may be provided with an interactive interface, for example, through which a user may input two entity words of which the association relationship is to be determined and select the estimated association relationship of the two entity words. Through the processing of terminal device 110, terminal device 110 may output a confidence that the two entity words have the estimated association.

According to the embodiment of the application, as shown in fig. 1, the application scenario may further include a server 130, and the terminal device 110 and the server 130 may communicate with each other through a network. For example, terminal device 110 may obtain a pre-trained process model from server 130 via a network.

Illustratively, the server 130 may also be, for example, various servers that provide support for applications running in the terminal device 110. The server 130 may also receive, for example, target association information sent by the terminal device according to two entity words input by the user and the selected estimated association relationship via the network, and determine, according to the target association information and the processing model, a confidence that the two entities represented by the two entity words have the estimated association relationship therebetween. For example, the server 130 may be an application server, a server of a distributed system, or a server incorporating a blockchain, for example. Alternatively, the server may also be a virtual server, a cloud server, or the like.

According to an embodiment of the present application, as shown in fig. 1, an application scenario of the embodiment may further include a first database 140 and a second database 150. The first database 140 stores a large number of text paragraphs (hereinafter, referred to as "text paragraphs"), and the second database 150 stores a pre-constructed Knowledge Graph (KG). The server 130 or the terminal device 110 may access the first database 140 and the second database 150 through the network to obtain the text segment similar to the target association information from the first database 140, and extract the associated entity word having an association relationship with the two entity words from the knowledge graph. It is understood that the first database 140 and the second database 150 may be two different databases or may be different memory partitions in the same database. In an embodiment, the first database 140 and the second database 150 may be two storage partitions in the server 130, for example, or either of the first database 140 and the second database 150 may be a storage disk or a cloud database.

According to the embodiment of the application, similar text segments and related entity words can be comprehensively considered, so that the characteristics of context information and the like of the two entity words in the text segments are extracted, and the confidence degree of the estimated association relationship between the two entities is determined according to the extracted characteristics.

According to the embodiment of the application, the terminal device 110 or the server 130 may further have a text recognition function, for example, to recognize the entity word and the association relationship between two entity words from the text. Confidence prediction can be carried out on the incidence relation between any two recognized entity words so as to verify the accuracy of the text recognition function.

It should be noted that the method for determining an association relationship between entities provided in the embodiment of the present application may be generally executed by the terminal device 110, or may also be executed by the server 130. Accordingly, the apparatus for determining the association relationship between the entities provided in the embodiment of the present application may be generally disposed in the terminal device 110, or may also be disposed in the server 130.

It should be understood that the types of terminal device, server, first database, and second database in fig. 1 are merely illustrative. There may be any type of terminal device, server, first database and second database, depending on implementation needs.

The method for monitoring the driving state provided by the embodiment of the present application is described in detail with reference to fig. 2 to 6 in the following application scenario described in fig. 1.

Fig. 2 is a flowchart illustrating a method for determining an association relationship between entities according to an embodiment of the present application.

As shown in fig. 2, the method 200 of determining an association relationship between entities of this embodiment may include operation S210, operation S230, operation S250, operation S270, and operation S290. The method 200 may be performed by the terminal device described above, or may be performed by a server, for example.

In operation S210, target association information is acquired.

According to an embodiment of the application, the target association information includes a first word representing the first entity, a second word representing the second entity, and a third word representing the target association relationship.

Illustratively, the first word, the second word and the third word may all be obtained in response to an input operation by a user, for example. The first word, the second word and the third word may form a Subject-prediction-Object (SPO) triple and use the triple as target association information. Where S, O indicate two words representing entities, respectively, and P indicates a third word representing an associative relationship between the two entities.

Illustratively, the first word and the second word may be extracted by performing entity recognition on the text. The third word may be a preset word, or may be obtained by learning a semantic relation between two entity words representing two entities in the text. After the first word and the second word are extracted from the text, a triple SPO may be formed with the third word, and the triple is used as the target association information.

According to an embodiment of the present application, the target association relationship represented by the third word may include any one of the following: upper and lower relationships, causal relationships, concomitant relationships, concurrency relationships, timing relationships, and the like. The target association relationship may be set according to an actual application scenario, which is not limited in the present application. For example, in the medical field, since the upper and lower relationships between entities representing diseases are expensive to construct, the target association relationship may be an upper and lower relationship in order to construct the upper and lower relationships of diseases.

In operation S230, a first semantic feature is determined according to the target associated information and a segment similar to the target associated information in a preset segment library.

According to the embodiment of the application, a large number of segments are stored in the preset segment library, and the large number of segments can be obtained by extracting text information acquired from various information propagation channels. The text information acquired by the information dissemination channel can include, for example, journal articles, professional books, newspaper advertisement text, news report text, and the like.

For example, the text information may be collected in advance and may be electronized by scanning or the like. The electronic text information is then recognized using Optical Character Recognition (OCR) technology or the like to segment the text information. Each segment in the preset segment library may record information of a segment in the text information. It will be appreciated that there may be different libraries of preset segments for different domains. In order to improve the accuracy and recall rate of the similar segments and improve the accuracy of the expression information of the similar segments, the segment library including a plurality of segments describing the professional knowledge in the field to which the first entity and the second entity belong may be used as the preset segment library.

For example, in order to facilitate the reference and the call, after obtaining a large number of segments, for example, an index may be created for the large number of segments, and the large number of segments may be sequentially stored in the first database described above according to the created index.

According to an embodiment of the present application, in operation S230, a segment similar to the target association information may be extracted from the preset segment library as a similar segment for the target association information. And then carrying out information fusion on the target associated information and the similar text segments to extract the characteristic of the context information representing the target associated information, and taking the characteristic as a first semantic characteristic.

For example, a Named Entity Recognition (NER) tool may be used to perform text Recognition on each segment in the preset segment library, so as to obtain Entity words included in each segment. And then extracting the segment of which the identified entity word comprises the first word and the second word from the plurality of segments as a similar segment. It is to be understood that the tools for text recognition described above are only examples to facilitate understanding of the present application and are not intended to be limiting.

Illustratively, a term-frequency-inverse-text-frequency (TF-IDF) statistical tool may be further used to process each segment in the segment library, and the statistical result is the importance degree of the first word and the second word in each segment. And then extracting the segments with the importance degree of the first word and the second word larger than the preset degree from the plurality of segments as similar segments.

According to the embodiment of the application, after the similar text segment is obtained, the target associated information and the similar text segment can be used as the input of the first semantic feature extraction model, and the first semantic feature is obtained after the processing of the semantic feature extraction model. The first semantic feature extraction model may be, for example, a Long Short-Term Memory network (LSTM) model, a semantic Representation model (ELMo), a Bidirectional Encoder (BERT), or the like.

According to the method and the device, the first semantic features are obtained by fusing the target associated information and the similar text segments, so that the extracted first semantic features can represent semantic information, lexical information, syntactic information and the like of the target associated information in the text segments, and the richness and accuracy of the extracted first semantic features are improved conveniently.

In operation S250, a description text for the target associated information is determined according to the target associated information and a preset knowledge graph.

According to an embodiment of the present application, the predetermined knowledge-graph may be constructed, for example, from a large amount of data in the field to which the first entity and the second entity belong. For example, entities, relationships, and entity attribute information may be first extracted from a large amount of data. And then taking the entities and the entity attributes as nodes, and performing edge connection on the nodes according to the relationship between the entities and the corresponding relationship between the entities and the entity attributes to obtain the knowledge graph.

Illustratively, when the domain to which the first entity and the second entity belong is a medical domain, the entities may include, for example, medical terms such as disease names, drug names, department names, etc., and the entity attributes may include, for example, clinical symptoms of the disease, symptoms targeted by the drug, etc. The relationship may include, for example, an attribution relationship between a disease and a treatment department, a companion relationship between symptoms, a concurrence relationship between diseases, and the like.

According to an embodiment of the present application, operation S250 may first determine two nodes in the knowledge-graph indicating two entities represented by the first word and the second word in the target associated information. And then extracting to obtain the associated nodes according to the nodes connected with the two nodes in the knowledge graph. And finally, splicing the entity words indicated by the associated nodes to form a description text.

For example, when extracting the associated node from the knowledge graph, a node indicating that the first word represents the entity may be used as the starting point, and a node having an association relationship with the starting point may be determined as the first node according to a connecting edge of the starting point. And then determining a node having a relationship with the first node as a second node according to the connecting edge of the first node by taking the first node as a starting point. And the like until the (m +1) th node which can establish the association relationship with the node indicating the first word representation entity through the m nodes is obtained. Finally, the first to (m +1) th nodes are used as the associated nodes. M is a natural number, and the value of m can be set according to actual requirements, which is not limited in the application.

In operation S270, a second semantic feature is determined according to the target association information and the description text.

According to an embodiment of the present application, in operation S270, the target related information and the description text may be used as inputs of a second semantic feature extraction model similar to the first semantic feature extraction model described above, and a second semantic feature vector is obtained through output. The first semantic feature extraction model and the second semantic feature extraction model are different in that training samples adopted in training are different.

Illustratively, the second semantic feature extraction model may be, for example, BERT, LSTM, word representation Global vector model (GloVe) or the like.

In operation S290, a confidence level that the first entity and the second entity have the target association relationship is determined according to the first semantic feature and the second semantic feature.

According to an embodiment of the present application, the operation S290 may, for example, use the first semantic feature and the second semantic feature as input of a prediction model at the same time, and obtain, via the prediction model, a confidence that two entities indicated by the first word and the second word have the target association relationship.

Illustratively, the predictive model may be, for example, a classifier for classification. The confidence vectors of the first word and the second word for the preset number of incidence relations can be output through the classifier. The confidence vector comprises a plurality of values, each value corresponds to one incidence relation in a preset number of incidence relations, and the value of each value is the confidence of the incidence relation between the two entities indicated by the first word and the second word. And then determining the value of the corresponding target incidence relation in the confidence coefficient vector as a target value. And taking the value of the target value as the finally determined confidence degree of the target association relationship between the first entity and the second entity.

Compared with the technical scheme of directly extracting the target association relationship from the text in the related technology, the method for determining the association relationship between the entities in the embodiment of the application extracts the similar text from the text library, obtains the description text based on the knowledge graph, finally determines the confidence coefficient of the relationship between the two entities by integrating the similar text and the description text, and can fully utilize the constructed structured information and the unstructured information. Therefore, the accuracy of the determined relation confidence coefficient can be effectively improved, the accurate electronic knowledge can be conveniently established, and the accuracy of information recall is favorably improved.

It is understood that the present application does not limit the execution sequence of the operations S230 and S250. For example, operation S230 and operation S250 may be performed simultaneously, or operation S230 may be performed before operation S250 is performed.

Fig. 3 is a schematic diagram of the principle of determining a first semantic feature according to an embodiment of the present application.

According to an embodiment of the present application, an Enhanced reconstruction through Knowledge Integration (Ernie) of continuous learning based on a fly-rotor-on source, which is capable of learning a complete semantic Representation, may be employed to determine the first semantic feature.

For example, when the first semantic feature is determined, for example, the target associated information may be split, a first word S representing the first entity and a third word P representing the target association relationship are converted into a question sentence through a template, and a second word O representing the second entity is replaced with a preset word, so as to obtain the target associated information of which the second word is masked by the preset word. And then, using target associated information of the shielded second word formed by splicing the question sentence and the preset word and the similar text segment described in the foregoing as the input of the pretrained Ernie model, and outputting the pretrained Ernie model to obtain the first semantic feature. The preset word may be, for example, any word indicating the type of the second word O, and may be, for example, "option," "option," or "lower-level word" or the like. The template may be set according to the target association relationship and the actual requirement, which is not limited in the present application. For example, the target association relationship represented by the third word is a top-bottom relationship, and the template may be "is the hyponym of a? ". When the first word is "ill", the question sentence converted from the template may be "ill hyponym? ".

According to the embodiment of the application, when the target associated information of the masked second word and the similar text segment are used as the input of the first semantic feature extraction model to obtain the first semantic feature, the dimension reduction processing can be further performed on the output vector of the Ernie model, and the vector after the dimension reduction processing is used as the first semantic feature. Therefore, the expression capability of the first semantic feature can be improved, and the accuracy of the confidence coefficient determined according to the first semantic feature can be improved conveniently.

For example, as shown in the embodiment 300 shown in fig. 3, when obtaining the first semantic feature, the target related information 310 may be preprocessed to replace the second word with the preset word, so as to obtain the masked information 320 of the masked second word. This occluded information 320 and the retrieved similar segments 330 are then used as input to a first semantic vector model 340(Ernie model). After being processed by the first semantic vector model 340, the embedded vector 350 is output. The embedded vector 350 is then used as an input to the attention neural network model 360 to perform a dimensionality reduction process on the embedded vector 350, and the vector output by the attention neural network model 360 is used as the first semantic feature 370.

Illustratively, the attention neural network model 360 may be, for example, a multi-layered cooperative attention (c)_o-attention). The attention neural network model 360 may be composed of, for example, a plurality of attention mechanism layers and a fully connected layer. The embedded vectors 350 are respectively used as inputs to the plurality of attention mechanism layers, and then outputs of the plurality of attention mechanism layers are simultaneously input to the fully-connected layer, and the first semantic features 370 are obtained through the fully-connected layer output. It can be understood that, the number of the power generation layers can be set according to actual requirements, which is not limited in the present application.

According to embodiments of the present application, the Ernie model may directly employ an existing pre-training model, for example. Or after the existing pre-training model is obtained, mass texts in the fields to which the first entity and the second entity belong may be obtained first, a word or an entity in the mass text is masked (mask) to obtain a training sample, and finally, the training sample is used to adjust model parameters in the existing training model to obtain a first semantic feature extraction model capable of being used for feature extraction.

FIG. 4 is a schematic diagram of determining a first semantic feature according to another embodiment of the present application.

According to the embodiment of the application, when the first semantic feature is determined, the text segments obtained from the preset text segment library according to the target associated information can be screened, and the text segments with high similarity to the target associated information are used as the similar text segments. By the method, the large noise caused by feature extraction of the text segment with low similarity can be avoided, and the accuracy of the extracted first semantic feature can be improved.

For example, as in the embodiment 400 shown in fig. 4, after the target related information is obtained, the first word 411, the second word 412, and the third word 413 in the target related information may be combined to obtain the search sentence. For example, a string obtained by concatenating the first word 411, the second word 412, and the third word 413 in a random order is used as the search term. A preset segment library 430 having a large number of segments is then retrieved according to the retrieval statement 420. The segment including all or part of the words in the search statement 420 is obtained from the preset segment library 430 as the candidate similar segment 440 for the target associated information. Then, the similarity between each candidate similar segment in the candidate similar segments 440 and the retrieval statement 420 is determined, and the segment whose similarity with the retrieval statement 420 satisfies the first preset condition is determined to be the similar segment 450 for the target association information. Finally, a first semantic vector 480 is obtained by using the similar segment 450 and the target associated information of the masked second word (i.e. the masked information 460) as the input of the first semantic feature extraction model 470.

Illustratively, when determining the similarity between each candidate similar segment and the search sentence 420, for example, a TF-IDF tool may be used to count the frequency that each candidate similar segment and the search sentence includes words in the word library, so as to convert each candidate similar segment and the search sentence into a word vector. And finally, taking the similarity between the two word vectors as the similarity between the alternative similar text segment and the retrieval sentence. The similarity can be embodied in any one of the following parameter forms: cosine similarity, Jacobs's similarity coefficient, Spireman correlation coefficient, etc.

Illustratively, a Binary Independent Model (BIM) based BM25 model may also be employed to compute the similarity between each candidate similar segment and the search statement 420. The BM25 model is a classic algorithm used in the field of information indexing to calculate similarity scores between query sentences and documents.

For example, the similarity satisfying the first preset condition may be a similarity greater than the first preset value, or a similarity ranked before the first predetermined position in order from the highest to the lowest. The first preset value and the first preset position can be set according to actual requirements, and the first preset value and the first preset position are not limited in the application. For example, the first preset value may be any value greater than 0.5, and the first predetermined position may be any position after the first bit and before the fifth bit.

Fig. 5 is a schematic diagram of a principle of determining a description text for target associated information according to an embodiment of the present application.

According to the embodiment of the application, the description text for the target associated information may be a text obtained by converting an SPO, for example. The embodiment may extract a plurality of SPO groups from the knowledge-graph as associated SPOs associated with the entities represented by the target association information based on the target association information. The associated SPO is then converted into a plurality of texts. And finally, extracting a text with higher similarity with the text converted by the target associated information from the plurality of texts converted according to the associated SPO as a description text for the target associated information. By extracting texts with high similarity from the plurality of texts as description texts and determining the second semantic features according to the description texts, the accuracy of the determined second semantic features in representing the target associated information can be improved, and the accuracy of the determined confidence coefficient is improved conveniently.

According to the embodiment of the present application, as in the embodiment 500 shown in fig. 5, when determining the description text for the target associated information, a first associated word having an association relationship with the first word 511 may be obtained from the preset knowledge graph 520, and a second associated word having an association relationship with the second word 512 may be obtained. A plurality of candidate associated information 540 is then determined based on the association relationship between the first word 511 and the first associated word and the association relationship between the second word 512 and the second associated word.

Illustratively, the first related word may include a word indicated by a first node in the preset knowledge graph 520 connected with the node indicating the first word 511 through an edge. The first related word may further include, for example, a word indicated by a second node connected to the first node through an edge, or the like. The first related word includes a word indicated by the related node determined according to m described above. After the first related word is obtained, a plurality of SPOs may be constructed and obtained according to edges connecting a plurality of nodes among the node indicating the first related word and the node indicating the first word. Each SPO in the plurality of SPOs is used as one piece of first associated information 531, and finally, a plurality of pieces of first associated information are obtained. It is to be understood that the second related word may be obtained and the second related information 532 may be obtained by a method similar to the method of obtaining the first related word. Finally, the first associated information 531 and the second associated information 532 are summarized to obtain a plurality of candidate associated information 540.

According to the embodiment of the application, after the alternative associated information is obtained, the description text of each alternative associated information and the description text of the target associated information can be determined according to the preset text template. And finally, determining a description text of which the similarity with the description text of the target associated information meets a second preset condition in the description texts of the multiple candidate associated information as the description text of the target associated information.

Illustratively, as shown in fig. 5, S, P, O in the target associated information 510 may be substituted into the text template 550 to obtain the target description text 560. And substituting S, P, O in each alternative associated information 540 into the text template 550 to obtain an alternative description text 570. The similarity between each alternative description text 570 and the target description text 560 is then calculated, resulting in a similarity for each alternative description text 570. Finally, the alternative description text for which the similarity satisfying the second preset condition is directed is taken as the finally determined description text 580 for the target associated information.

For example, the similarity satisfying the second preset condition may be a similarity greater than the second preset value, or a similarity ranked before the second predetermined position in order from the highest to the lowest. The second preset value and the second preset position can be set according to actual requirements, and the method is not limited in the application. For example, the second preset value may be any value greater than 0.6, and the second predetermined position may be any position after the second bit and before the eighth bit. It is understood that the second preset condition may be the same as the first preset condition described above, for example, according to actual requirements.

For example, the preset text template may be set according to actual requirements, for example, the preset text template may be "a has C association with B". When the target associated information is 'sick-upper-lower relation-cold', the description text 'sick-upper-lower relation with cold' can be obtained by substituting the target associated information into the preset text template. It is to be understood that the preset text template is only used as an example to facilitate the understanding of the present application, and the present application is not limited thereto.

According to the embodiment of the application, after the description text for the target associated information is obtained, the operation of replacing the second word with the preset word as described above may be performed on the target associated information to obtain the target associated information of the masked second word. And finally, the target associated information of the shielded second word and the description text aiming at the target associated information are used as the input of a second semantic feature extraction model, and the second semantic feature is obtained through the output after the processing of the second semantic feature extraction model. It will be appreciated that the second semantic features may be derived in a similar way to the method of deriving the first semantic features, the difference between the two methods being that the text on which the second semantic features are derived is descriptive text of structured knowledge extracted from the knowledge-graph.

According to the embodiment of the application, after the first semantic feature and the second semantic feature are obtained, the first semantic feature and the second semantic feature can be simultaneously input into the prediction model, and a probability vector is output after the prediction model is processed, wherein the probability vector comprises probability values of the masked second word as each entity word in a preset entity word bank. Finally, a probability value of the probability vector indicating that the masked second word is an entity word of the second entity may be determined as a confidence level that the first entity and the second entity have the target association relationship.

Illustratively, the predetermined entity thesaurus may include entity words extracted from expertise describing areas to which the first entity and the second entity belong, the entity thesaurus including the second word. For different fields, different entity word banks can be constructed in advance. The determined probability value is a probability value of the probability vector for an entity word describing the second entity in the preset entity word bank.

Illustratively, the predictive model may be, for example, a classification model, which may include, for example, a fully connected layer and an output layer. The full connection layer is used for fusing the first semantic features and the second semantic features, and the output layer is used for outputting the probability vectors described in the foregoing. Illustratively, the classification model may further include a softmax activation layer for normalizing the output probability vector.

Fig. 6 is a schematic diagram illustrating a principle of determining an association relationship between entities according to an embodiment of the present application.

According to an embodiment of the present application, as shown in fig. 6, in this embodiment 600, the overall process of determining the association relationship between the entities may include a process described below.

After the target associated information is obtained, the second word is used as an option 612, the first word and the third word are converted into a question 613 according to a template, the first word, the second word and the third word are spliced to form a retrieval statement, and a similar text segment 611 for the target associated information is obtained through retrieval. Segment 611 is then used as evidence, the evidence, option 612 and question 613 are spliced and input into first semantic feature extraction model 620, and the result is processed by first semantic feature extraction model 620 and output to obtain an embedded vector capable of expressing segments, options and questions, wherein the embedded vector can be understood as a vector obtained by splicing segment expression vector 631, option expression vector 632 and question expression vector 633. Then, the embedded vector output by the first semantic feature extraction model 620 is input into the attention neural network model 640, and the first semantic feature is output.

Meanwhile, a relevant word for the first word and a relevant word for the second word may be obtained from the knowledge graph 650, so as to obtain a plurality of combinations of S and O, where S and O are two words respectively indicated by two nodes having a connecting edge in the knowledge graph. According to the association relationship indicated by the S and O and the edges connecting between the two nodes, a plurality of SPOs can be obtained, so that the association character string 662 is formed by splicing. Subsequently, the associated character string 662 is converted into a description text using a preset text template. Then, the target related information of the masked second word obtained by masking the second word is used as a target character string 661 and a description text to be input into the second semantic feature extraction model 670, and the second semantic feature is output after being processed by the second semantic feature extraction model 670.

Finally, the first semantic feature and the second semantic feature are simultaneously input into the prediction model 680, and the confidence 690 that the first entity and the second entity have the target association relationship can be obtained according to the output of the prediction model 680.

According to the method for determining the association relationship between the entities, the existing structured knowledge and unstructured knowledge can be fully utilized to determine the confidence coefficient of the association relationship, and therefore the method can be used for recalling the association relationship between the entities. Compared with the prior art that the incidence relation is recalled only according to the text, the accuracy of the recalled incidence relation can be improved. By applying the method to the prediction of the upper-lower relation between the entities in the medical field, the construction cost of the upper-lower relation of the diseases can be effectively reduced, and the intelligent electronic management of knowledge in the medical field is facilitated.

According to the embodiment of the application, after the confidence degree of the target association relation between the first entity and the second entity is obtained, if the confidence degree is higher, the knowledge graph can be supplemented according to the target association information, so that more comprehensive and richer structural knowledge can be obtained in subsequent confidence degree prediction, and the accuracy of the subsequent confidence degree prediction is improved.

According to an embodiment of the present application, the operation of determining the association relationship between the entities in this embodiment may further determine whether the confidence of the target association relationship between the first entity and the second entity is greater than a preset confidence. And if the target association information is greater than the preset confidence level, supplementing the preset knowledge graph according to the target association information.

For example, when supplementing the preset knowledge-graph, it may be determined whether the preset knowledge-graph has a node indicating the first word and a node indicating the second word. And if the node indicating the first word is not included, adding a node in the knowledge graph for indicating the first word. Similarly, a node indicating the second word may be added when a node indicating the second word is not included. After determining that the preset knowledge graph has the node indicating the first word and the node indicating the second word, determining whether a connecting edge exists between the node indicating the first word and the node indicating the second word, and if not, adding the edge between the two nodes to complete the supplement of the preset knowledge graph.

For example, the preset confidence level may be determined according to the prediction confidence level obtained by using the test sample after training each model. For example, the confidence level may be set according to a relationship between the confidence level and the label of the test sample, so that the preset confidence level is set to a value that enables the confidence level of the test sample for a preset proportion to be greater than the preset confidence level. The preset ratio may be a large value such as 80%, 85%, or 90%, and the preset confidence may be a value greater than 0.5, such as 0.6 or 0.8. It is to be understood that the values and determination methods of the preset ratio and the preset confidence are only used as examples to facilitate understanding of the present application, and the present application is not limited thereto.

According to the embodiment of the application, after determining that the confidence level of the target association relationship between the first entity and the second entity is greater than the preset confidence level, for example, a professional person may perform confidence level accuracy evaluation. And when the evaluation result is that the confidence degree is credible, supplementing a preset knowledge graph according to the target association relation.

Based on the method for determining the association relationship between the entities described above, the present application also provides a device for determining the association relationship between the entities. The apparatus provided in the present application will be described in detail below with reference to fig. 7.

Fig. 7 is a block diagram of an apparatus for determining an association relationship between entities according to an embodiment of the present application.

As shown in fig. 7, the apparatus 700 for determining an association relationship between entities of this embodiment may include an information obtaining module 710, a first feature determining module 730, a text determining module 750, a second feature determining module 770, and a confidence determining module 790.

The information obtaining module 710 is configured to obtain target associated information, where the target associated information includes a first word representing a first entity, a second word representing a second entity, and a third word representing a target association relationship. In an embodiment, the information obtaining module 710 may be configured to perform the operation S210 described above, for example, and is not described herein again.

The first feature determining module 730 is configured to determine a first semantic feature according to the target associated information and a segment similar to the target associated information in a preset segment library. In an embodiment, the first characteristic determining module 730 may be configured to perform the operation S230 described above, for example, and is not described herein again.

The text determining module 750 is configured to determine a description text for the target associated information according to the target associated information and a preset knowledge graph. In an embodiment, the text determining module 750 may be configured to perform the operation S250 described above, for example, and is not described herein again.

The second feature determining module 770 is configured to determine a second semantic feature according to the target related information and the description text. In an embodiment, the second characteristic determining module 770 may be configured to perform the operation S270 described above, for example, and will not be described herein again.

The confidence determination module 790 is configured to determine a confidence that the first entity and the second entity have the target association relationship according to the first semantic feature and the second semantic feature. In an embodiment, the confidence determining module 790 may be configured to perform the operation S290 described above, for example, and will not be described herein again.

According to an embodiment of the present application, an electronic device and a readable storage medium are also provided.

Fig. 8 is a block diagram of an electronic device for implementing a method for determining an association relationship between entities according to an embodiment of the present application.

Electronic devices are intended to represent various forms of digital computers, such as, for example, an oxazine computer, a desktop computer, a workstation, a personal digital assistant, a server, a blade server, a mainframe computer, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital processing, cellular phones, smart phones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions, are meant to be examples only, and are not meant to limit implementations of the present application that are described and/or claimed herein.

As shown in fig. 8, the electronic device 800 includes: one or more processors 801, memory 802, and interfaces for connecting the various components, including a high speed interface and a low speed interface. The various components are interconnected using different buses and may be mounted on a common motherboard or in other manners as desired. The processor may process instructions for execution within the electronic device, including instructions stored in or on the memory to display graphical information of a GUI on an external input/output apparatus (such as a display device coupled to the interface). In other embodiments, multiple processors and/or multiple buses may be used, along with multiple memories and multiple memories, as desired. Also, multiple electronic devices may be connected, with each device providing portions of the necessary operations (e.g., as a server array, a group of blade servers, or a multi-processor system). Fig. 8 illustrates an example of a processor 801.

The memory 802 is a non-transitory computer readable storage medium as provided herein. Wherein the memory stores instructions executable by at least one processor to cause the at least one processor to perform the method of determining associations between entities as provided herein. The non-transitory computer readable storage medium of the present application stores computer instructions for causing a computer to perform the method of determining an association between entities provided herein.

The memory 802, which is a non-transitory computer-readable storage medium, may be used to store non-transitory software programs, non-transitory computer-executable programs, and modules, such as program instructions/modules corresponding to the method of determining associations between entities in embodiments of the present application (e.g., the information acquisition module 710, the first feature determination module 730, the text determination module 750, the second feature determination module 770, and the confidence determination module 790 shown in fig. 7). The processor 801 executes various functional applications of the server and data processing by running non-transitory software programs, instructions, and modules stored in the memory 802, that is, implements the method of determining the association relationship between entities in the above-described method embodiments.

The memory 802 may include a storage program area and a storage data area, wherein the storage program area may store an operating system, an application program required for at least one function; the storage data area may store data created according to use of an electronic device implementing the method of determining an association between entities, and the like. Further, the memory 802 may include high speed random access memory and may also include non-transitory memory, such as at least one magnetic disk storage device, flash memory device, or other non-transitory solid state storage device. In some embodiments, the memory 802 optionally includes memory located remotely from the processor 801, which may be connected via a network to an electronic device implementing the method of determining associations between entities. Examples of such networks include, but are not limited to, the internet, intranets, local area networks, mobile communication networks, and combinations thereof.

The electronic device 800 implementing the method of determining an association between entities may further comprise: an input device 803 and an output device 804. The processor 801, the memory 802, the input device 803, and the output device 804 may be connected by a bus or other means, and are exemplified by a bus in fig. 8.

The input device 803 may receive input numeric or character information and generate key signal inputs related to user settings and function controls of an electronic apparatus implementing the method of determining an association between entities, such as a touch screen, a keypad, a mouse, a track pad, a touch pad, a pointer, one or more mouse buttons, a track ball, a joystick, or other input devices. The output devices 804 may include a display device, auxiliary lighting devices (e.g., LEDs), and haptic feedback devices (e.g., vibrating motors), among others. The display device may include, but is not limited to, a Liquid Crystal Display (LCD), a Light Emitting Diode (LED) display, and a plasma display. In some implementations, the display device can be a touch screen.

Various implementations of the systems and techniques described here can be realized in digital electronic circuitry, integrated circuitry, application specific ASICs (application specific integrated circuits), computer hardware, firmware, software, and/or combinations thereof. These various embodiments may include: implemented in one or more computer programs that are executable and/or interpretable on a programmable system including at least one programmable processor, which may be special or general purpose, receiving data and instructions from, and transmitting data and instructions to, a storage system, at least one input device, and at least one output device.

These computer programs (also known as programs, software applications, or code) include machine instructions for a programmable processor, and may be implemented using high-level procedural and/or object-oriented programming languages, and/or assembly/machine languages. As used herein, the terms "machine-readable medium" and "computer-readable medium" refer to any computer program product, apparatus, and/or device (e.g., magnetic discs, optical disks, memory, Programmable Logic Devices (PLDs)) used to provide machine instructions and/or data to a programmable processor, including a machine-readable medium that receives machine instructions as a machine-readable signal. The term "machine-readable signal" refers to any signal used to provide machine instructions and/or data to a programmable processor.

To provide for interaction with a user, the systems and techniques described here can be implemented on a computer having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to a user; and a keyboard and a pointing device (e.g., a mouse or a trackball) by which a user can provide input to the computer. Other kinds of devices may also be used to provide for interaction with a user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user may be received in any form, including acoustic, speech, or tactile input.

The systems and techniques described here can be implemented in a computing system that includes a back-end component (e.g., as a data server), or that includes a middleware component (e.g., an application server), or that includes a front-end component (e.g., a user computer having a graphical user interface or a web browser through which a user can interact with an implementation of the systems and techniques described here), or any combination of such back-end, middleware, or front-end components. The components of the system can be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include: local Area Networks (LANs), Wide Area Networks (WANs), and the Internet.

The computer system may include clients and servers. A client and server are generally remote from each other and typically interact through a communication network. The relationship of client and server arises by virtue of computer programs running on the respective computers and having a client-server relationship to each other.

It should be understood that various forms of the flows shown above may be used, with steps reordered, added, or deleted. For example, the steps described in the present application may be executed in parallel, sequentially, or in different orders, and the present invention is not limited thereto as long as the desired results of the technical solutions disclosed in the present application can be achieved.

The above-described embodiments should not be construed as limiting the scope of the present application. It should be understood by those skilled in the art that various modifications, combinations, sub-combinations and substitutions may be made in accordance with design requirements and other factors. Any modification, equivalent replacement, and improvement made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method of determining an associative relationship between entities, comprising:

acquiring target associated information, wherein the target associated information comprises a first word representing a first entity, a second word representing a second entity and a third word representing a target association relation;

determining a first semantic feature according to the target associated information and similar segments aiming at the target associated information in a preset segment library;

determining a description text aiming at the target associated information according to the target associated information and a preset knowledge graph;

determining a second semantic feature according to the target associated information and the description text; and

determining a confidence level that the first entity and the second entity have the target association relationship based on the first semantic feature and the second semantic feature.

2. The method of claim 1, wherein determining the first semantic feature comprises:

and taking the target associated information for shielding the second word and the similar text segment as the input of a first semantic feature extraction model to obtain the first semantic feature.

3. The method of claim 2, wherein the obtaining the first semantic feature by using the target associated information of the occluded second word and the similar segment as input of a first semantic feature extraction model comprises:

taking the target associated information of the second word and the similar text segment which are shielded as the input of a first semantic feature extraction model, and outputting an embedded vector; and

and outputting the first semantic features by taking the embedded vector as an input of an attention neural network model.

4. The method of any of claims 2-3, wherein determining the first semantic feature further comprises:

combining the first word, the second word and the third word to obtain a retrieval sentence;

retrieving the preset text segment library according to the retrieval statement to obtain alternative similar text segments aiming at the target associated information; and

and determining the text segment of which the similarity with the retrieval sentence in the alternative similar text segments meets a first preset condition as the similar text segment aiming at the target associated information.

5. The method of claim 1, wherein determining the descriptive text for the target associated information comprises:

acquiring a first associated word having an association relation with the first word and a second associated word having an association relation with the second word from the preset knowledge graph;

determining a plurality of candidate associated information according to the association relationship between the first word and the first associated word and the association relationship between the second word and the second associated word;

determining a description text of the target associated information and a description text of each of the plurality of candidate associated information according to a preset text template; and

and determining a description text of which the similarity with the description text of the target associated information meets a second preset condition in the description texts of the multiple candidate associated information as the description text of the target associated information.

6. The method of claim 1, wherein determining a second semantic feature comprises:

and outputting to obtain the second semantic features by taking the target associated information for shielding the second words and the description text aiming at the target associated information as the input of a second semantic feature extraction model.

7. The method of claim 1, wherein determining a confidence level that the first entity and the second entity have the target association comprises:

taking the first semantic features and the second semantic features as input of a prediction model, and outputting the shielded second words as probability vectors of all entity words in a preset entity word bank; and

determining a probability of the probability vector indicating the occluded second word as an entity word of the second entity as a confidence level of the target association relationship between the first entity and the second entity.

8. The method of claim 1, further comprising:

and supplementing the preset knowledge graph according to the target association information under the condition that the confidence coefficient of the target association relation between the first entity and the second entity is greater than a preset confidence coefficient.

9. The method of claim 1, wherein the library of preset segments comprises: a plurality of segments describing expertise in which the first entity and the second entity belong.

10. An apparatus for determining an associative relationship between entities, comprising:

the information acquisition module is used for acquiring target associated information, wherein the target associated information comprises a first word representing a first entity, a second word representing a second entity and a third word representing a target association relation;

the first characteristic determining module is used for determining a first semantic characteristic according to the target associated information and similar segments aiming at the target associated information in a preset segment library;

the text determination module is used for determining a description text aiming at the target associated information according to the target associated information and a preset knowledge graph;

the second characteristic determining module is used for determining a second semantic characteristic according to the target associated information and the description text; and

a confidence determination module, configured to determine a confidence that the target association relationship exists between the first entity and the second entity according to the first semantic feature and the second semantic feature.

11. An electronic device, comprising:

at least one processor; and

a memory communicatively coupled to the at least one processor; wherein the content of the first and second substances,

the memory stores instructions executable by the at least one processor to cause the at least one processor to perform: the method of any one of claims 1 to 9.

12. A non-transitory computer readable storage medium storing computer instructions for causing a computer to perform: the method of any one of claims 1 to 9.