Method and device for presenting hint information, and computer program

Abstract

To provide a method and device for presenting hint information and a computer program.SOLUTION: A method includes: inputting an entity, a context of the entity, a type of the entity, a part of speech of the context, and multiple predefined concepts; generating a first vector based on a combination of the entity, the context, the type of the entity and the part of speech of the context; generating a second vector based on each of the multiple concepts; generating a third vector based on a graph; generating a fourth vector by concatenating the second vector and the third vector; calculating a semantic similarity between the entity and each of the multiple concepts based on the first vector and the fourth vector; determining, based on the first vector and the semantic similarity, one concept corresponding the entity; and generating hint information to be presented to a user based on the determined concept.SELECTED DRAWING: Figure 4

Description

[Technical Field]

[0001] The present invention relates to entity disambiguation techniques, and more particularly to a method, apparatus, and computer program for presenting hint information about entities in text to a user using entity disambiguation techniques. [Background technology]

[0002] In real language environments, there are often situations where one entity name corresponds to multiple concepts. For example, the entity name "apple" that appears in a text may refer to both the fruit and the company Apple Inc. To resolve the ambiguity caused by the same entity name, entity disambiguation techniques have been proposed.

[0003] Entity disambiguation techniques can link contextual mentions, or entities, to appropriate concepts in a knowledge graph, which plays a fundamental role in many fields, such as question answering, semantic search, and information extraction. A concept is a distinct, independently existing entity, and there are interrelationships between different concepts in a knowledge graph. For example, two people with similar names exist: "Michael Jeffrey Jordan" and "Michael Owen Jordan." The former is a basketball star, and the latter is a well-known figure in the field of machine learning. Thus, a knowledge graph might contain two concepts: "Michael Jeffrey Jordan, basketball star" and "Michael Owen Jordan, machine learning expert." Furthermore, a knowledge graph might contain multiple sports concepts associated with "Michael Jeffrey Jordan, basketball star," and multiple computer technology concepts associated with "Michael Owen Jordan, machine learning expert." When a piece of text contains the entity "Michael Jordan," it is necessary to determine, based on the context of the text, whether this entity refers to "Michael Jeffrey Jordan, basketball star" or "Michael Owen Jordan, machine learning expert" in the Knowledge Graph.

[0004] After using entity disambiguation technology to determine that an entity mentioned in a text corresponds to a specific concept, hint information can be provided to the user based on the determined concept, allowing the user to understand the correct meaning of the entity. For example, for the entity "Apple" that appears in the text "Apple released a new smartphone today...", hint information such as "Apple: a company in the United States, a high-tech company, whose representative product is the iPhone (registered trademark) smartphone..." can be provided to the user.

[0005] However, due to the high complexity of natural language, entity disambiguation techniques face the challenge of how to identify the precise meaning of an entity in context and associate it with the correct concept in the knowledge graph.

[0006] Currently, there are two main types of entity disambiguation approaches: those that use reordering models to model the entity disambiguation problem, and those that utilize classification models to model the entity disambiguation problem.

[0007] The method using the reordering model includes two steps: generating candidate concepts and reordering the candidate concepts. In the step of generating candidate concepts, simple rules are often used, but this often leads to the inability to select accurate candidate concepts, which may result in cascading errors in the subsequent reordering step.

[0008] Methods using classification models model the entity disambiguation problem as a single-label text classification task. For example, the model described in the paper "Medical concept normalization in social media posts with recurrent neural networks" (Journal of Biomedical Informatics) by Elena Tutubalina et al., published in June 2018, includes two parts: a neural network and auxiliary features. The neural network part employs a gated recurrent unit (GRU) network and an attention mechanism network to encode entities. The auxiliary features part enhances the model using TF-IDF similarity and Word2Vec similarity. However, such models cannot accurately compare the semantic similarity between entities and concepts because the Word2Vec similarity feature lacks sufficient semantic information. Summary of the Invention [Problem to be solved by the invention]

[0009] To solve one or more problems in the prior art, the present invention proposes a new method using a classification model, which employs the BERT model to encode entities and employs character similarity and semantic similarity as auxiliary features. The method according to the present invention can increase the probability of associating entities mentioned in text with accurate concepts in a knowledge graph, thereby providing users with more accurate hint information. [Means for solving the problem]

[0010] According to one aspect of the present invention, there is provided a computer-implemented method for providing hint information to a user viewing an electronic text using a neural network, wherein the neural network includes a BERT model and a graph convolutional neural network, and the method includes: inputting the electronic text, information about the electronic text, and a plurality of predefined concepts into the neural network, wherein the electronic text includes entities and contexts of the entities, the information about the electronic text includes types of the entities and parts of speech of the contexts, and the concepts are in a text format; using the BERT model to generate a first vector (V2) based on a combination of the entity, the context, the type of the entity, and the part of speech of the context; generating a second vector (V4) based on each concept of the plurality of concepts using the BERT model; Using the graph convolutional neural network to generate a third vector (V5) based on a graph, wherein the graph is generated based on the plurality of concepts and the relationships between the plurality of concepts; generating a fourth vector (V6) by concatenating the second vector (V4) and the third vector (V5); Calculating a semantic similarity between the entity and each concept among the plurality of concepts based on the first vector (V2) and the fourth vector (V6); Determine one concept among the plurality of concepts corresponding to the entity based on the first vector (V2) and the semantic similarity; and The hint information is generated based on the determined concept corresponding to the entity and presented to the user.

[0011] According to another aspect of the present invention, there is provided an apparatus for providing hint information to a user viewing an electronic text using a neural network, wherein the neural network includes a BERT model and a graph convolutional neural network, the apparatus including: a memory for storing a computer program; and one or more processors connected to the memory, the processors executing the computer program to perform the following operations: inputting the electronic text, information about the electronic text, and a plurality of predefined concepts into the neural network, wherein the electronic text includes entities and contexts of the entities, the information about the electronic text includes types of the entities and parts of speech of the contexts, and the concepts are in text format; using the BERT model to generate a first vector (V2) based on a combination of the entity, the context, the type of the entity, and the part of speech of the context; generating a second vector (V4) based on each concept of the plurality of concepts using the BERT model; Using the graph convolutional neural network to generate a third vector (V5) based on a graph, wherein the graph is generated based on the plurality of concepts and the relationships between the plurality of concepts; generating a fourth vector (V6) by concatenating the second vector (V4) and the third vector (V5); Calculating a semantic similarity between the entity and each concept among the plurality of concepts based on the first vector (V2) and the fourth vector (V6); Determine one concept among the plurality of concepts corresponding to the entity based on the first vector (V2) and the semantic similarity; and This is an operation of generating the hint information based on the determined concept corresponding to the entity and presenting it to the user.

[0012] According to another aspect of the present invention, there is provided an apparatus for presenting hint information to a user browsing an electronic text, wherein the electronic text includes entities and contexts of the entities, the apparatus including: a main module, a semantic similarity calculation module, a character similarity calculation module, and a presentation module; The main module comprises: a BERT module configured to generate a first vector (V2) based on a combination of the entity, the context, the type of the entity, and the part of speech of the context; and a classification module configured to determine one concept among a plurality of predefined concepts corresponding to the entity based on the first vector (V2), the semantic similarity calculated by the semantic similarity calculation module, and the character similarity calculated by the character similarity calculation module, wherein the concept is in a text format; The semantic similarity calculation module a BERT module configured to generate a second vector (V4) based on each concept of the plurality of concepts; and a graph convolutional neural network configured to generate a third vector (V5) based on a graph, wherein the graph comprises a graph convolutional neural network generated based on the plurality of concepts and relationships between the plurality of concepts; the semantic similarity calculation module is configured to generate a fourth vector (V6) by concatenating the second vector (V4) and the third vector (V5), and calculate the semantic similarity between the entity and each concept among the plurality of concepts based on the first vector (V2) and the fourth vector (V6); the character similarity calculation module is configured to calculate a character similarity between the entity and each concept among the plurality of concepts; The presentation module is configured to generate the hint information based on the determined concept corresponding to the entity and present it to the user.

[0013] According to another aspect of the invention, there is provided a storage medium storing a computer program which, when executed by a computer, causes the computer to carry out a method as described above. [Brief explanation of the drawings]

[0014] [Figure 1] FIG. 1 is a diagram illustrating the architecture of a neural network according to the present invention. [Figure 2] FIG. 10 is a diagram showing combinations of entities, contexts, entity types, and context parts of speech. [Figure 3] This is a graph generated based on all concepts. [Figure 4] FIG. 1 is a diagram illustrating the configuration of an apparatus for presenting hint information about an entity to a user in an embodiment of the present invention. [Figure 5] 1 is a flowchart of a method for presenting hint information about an entity to a user in an embodiment of the present invention. [Figure 6] FIG. 1 illustrates an exemplary configuration of a general-purpose computer upon which embodiments of the present invention may be implemented. DETAILED DESCRIPTION OF THE INVENTION

[0015] DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings. However, these embodiments are merely illustrative and are not intended to limit the present invention.

[0016] FIG. 1 illustrates the architecture of a neural network in the present invention. The left half of FIG. 1 shows the main model, and the right half shows the auxiliary features. The main model is primarily used to perform classification tasks, i.e., to determine which concept (class) an entity contained in text belongs to. The main model employs a conventional Bidirectional Encoder Representation from Transformers (BERT) model. In neural network-based natural language processing, each word in a text is typically represented by a one-dimensional vector (also called a "word vector"). The BERT model uses the word vector as input and outputs a vector representation corresponding to the word that integrates (combines) the semantic information of the entire text. In this way, the BERT model can generate a semantic representation of a given text that contains rich semantic information. The generated semantic representation of the text is then fine-tuned for a specific natural language processing task, making it applicable to the specific task.

[0017] The main model and auxiliary features of the present invention will be described below in order.

[0018] For an entity that appears in a text, the present invention takes the combination of the entity, the context, the type of the entity, and the part of speech of the context as the input of the main model. Figure 2 shows an example of a specific form of such a combination.

[0019] As shown in Figure 2, the first line represents the text content. An opening flag [E] is added before the first character of the entity, and an closing flag [E / ] is added after the last character. Then, the entity's preamble is placed before the opening flag [E], and the entity's postamble is placed after the postamble [E / ]. In this way, the entity is matched with its context. Therefore, "M1 M2..." represents the entity string, and "L1 L2..." and "R1 R2..." represent the entity's preamble and postamble strings, respectively.

[0020] Furthermore, the dictionary resource is referenced to identify the entity type of the entity and the part of speech information of the corresponding context. The dictionary resource includes a predefined part of speech dictionary and an entity type dictionary. The part of speech dictionary defines parts of speech such as nouns, verbs, adverbs, and adjectives, and the entity type dictionary defines multiple entity types such as animals, companies, and games. The second line in Figure 2 shows the entity type of the entity in the main text determined by referencing the dictionary resource. x and context part of speech POS i This shows:

[0021] The text content in the first line in Figure 2 can be represented as a two-dimensional vector [batch_size, document_token], where batch_size indicates the number of input documents and document_token indicates the character ID sequence in each document. Similarly, the part of speech and entity type information in the second line can also be represented as a two-dimensional vector [batch_size, document_token].

[0022] Referring back to Figure 1, the Embedding and Adding layer 110 converts the input combination of entity, context, entity type, and context part of speech into a non-sparse vector, which is then input to the BERT model 120.

[0023] Specifically, the embedding and addition layer 110 converts the two-dimensional vector [batch_size,document_token] corresponding to the first row in Figure 2 into [batch_size,document_token,embedding_dim], where embedding_dim represents the embedding dimension. Similarly, the embedding and addition layer 110 converts the two-dimensional vector [batch_size,document_token] corresponding to the second row in Figure 2 into [batch_size,document_token,embedding_dim].

[0024] Then, based on the following formula (1), the embedding and addition layer 110 obtains the output vector V1 by calculating the weighted sum of the first and second rows.

[0025] V1=sigmoid(W1X1+W2X2) (1) Here, sigmoid denotes the activation function sigmoid = 1 / (1 + exp(-x)), W1 and W2 denote the weights waiting for training, and X1 and X2 denote the transformed vectors corresponding to the first and second rows in Figure 2, respectively.

[0026] The BERT model 120 receives the vector V1 output by the embedding and addition layer 110, and performs encoding on entities in the text content to extract semantic information of the entities. The BERT model is well known to those skilled in the art, so a detailed description thereof will be omitted in the present invention. As described above, the BERT model 120 receives the vector V1[batch_size,document_token,embedding_dim] as input, and the BERT model 120 outputs the vector V2[batch_size,bert_dim], where bert_dim represents the dimension of the hidden layer of the BERT model 120, i.e., the output dimension.

[0027] The splicing layer 130 receives the output V2 of the BERT model 120 and splices it with auxiliary features, which are described below. After splicing, the splicing layer 130 generates a vector V3[batch_size,bert_dim+2 * class_dim], where class_dim represents the number of categories. Since categories correspond to concepts in the knowledge graph, class_dim also represents the number of predefined concepts in the knowledge graph.

[0028] The output vector V3 of the splicing layer 130 is then input to the classification layer 140. As an example, the classification layer 140 can be implemented by a Softmax classifier. In the classification layer 140, the vector V3 is first converted into one-hot encoding format (one-hot encoding), and the converted dimensions are [batch_size, class_dim]. The classification layer 140 then generates a classification prediction result based on the converted vector. The prediction result indicates the probability that an entity in the text belongs to each concept (category) in the knowledge graph, and the concept corresponding to the maximum probability can be determined as the concept to which the entity belongs.

[0029] In particular, in practice, knowledge graphs usually cannot cover all existing concepts and only contain a limited number of concepts. Therefore, the following situation may occur: an entity does not belong to any concept in the knowledge graph, and using any concept to interpret the meaning of the entity is inappropriate. In such a case, a threshold may be set. When the maximum probability in the prediction result is greater than the threshold, the concept corresponding to the maximum probability is determined to be the concept to which the entity belongs, and hint information is generated for the user based on the content of the determined concept, thereby helping the user to accurately grasp the meaning of the entity. On the other hand, when all probabilities in the prediction result are smaller than the threshold, it means that it is inappropriate to classify the entity into any concept, and in this case, hint information for the entity is not generated.

[0030] The auxiliary features are described in detail below. In the present invention, similarity features between entities and concepts are calculated as auxiliary features to enhance the performance of the model. Preferably, to reduce the amount of calculation, only concepts in a predetermined knowledge graph are selected. The similarity features include character similarity and semantic similarity. As an example, in the present invention, BM25 similarity is used as the character similarity and vector similarity based on a neural network is used as the semantic similarity.

[0031] Specifically, when BM25 similarity is used as character similarity, the character similarity is calculated as follows: for an entity in the text, calculate the BM25 similarity between it and each concept (in the form of text) in the knowledge graph. As a result of the calculation, a vector with dimensions [batch_size, class_dim] is obtained. Note that since the BM25 algorithm (Best Match 25) is a text similarity algorithm well known to those skilled in the art, a detailed description of the calculation of BM25 similarity is omitted here.

[0032] Regarding semantic similarity, the present invention calculates the cosine similarity between the vector representation of the entity and the vector representation of each concept in the knowledge graph. As a result of the calculation, a vector with dimensions [batch_size, class_dim] is obtained. The method for calculating semantic similarity is described in detail below.

[0033] Referring to the auxiliary features shown in the right half of Figure 1, each concept in the knowledge graph is processed by the embedding and addition layer 110 and the BERT model 120, and then a vector V4 representing the semantic features of each concept is output from the BERT model 120. As shown in Figure 1, the main model and auxiliary features share the embedding and addition layer 110 and the BERT model 120.

[0034] On the other hand, a graph is constructed (generated) based on all concepts in the knowledge graph, and Figure 3 shows an example of the constructed graph. In Figure 3, each node N represents one concept, and the connecting lines between the nodes represent the interrelationships between concepts.

[0035] A Laplacian matrix is ​​calculated for the constructed graph, and then the Laplacian matrix is ​​input into a graph convolutional neural network (GCN) 150. Since the graph convolutional neural network (GCN) is a technology known to those skilled in the art, a detailed description thereof is omitted in this specification. The GCN 150 can output a vector of global features representing all concepts, which includes the interrelationships between each concept. Since concepts are associated with each other, the GCN 150 can model the interrelationships between concepts, as shown in FIG. 3. Therefore, the vector V5 generated by the GCN 150 includes the interrelationships between different concepts. The dimension of the vector V5 is [batch_size, graph_dim], where graph_dim represents the dimension of the output layer of the GCN 150.

[0036] Then, the output vector V4 of the BERT model 120 and the output vector V5 of the GCN 150 are all input to the splicing and dimension transformation layer 160. The splicing and dimension transformation layer 160 splices and transforms the dimensions of the vectors V4 and V5, and outputs a vector V6 with dimensions [batch_size,bert_dim]. Note that vector V6 is a vector representation of a concept in the knowledge graph.

[0037] Meanwhile, as described above, in the main model in the left half of Fig. 1, a vector representation V2 of an input entity is obtained by the BERT model 120. Then, the main model calculates the semantic similarity between the entity and the concept based on the vector representation V2 of the entity output by the BERT model 120 and the vector representation V6 of the concept output by the splicing and dimension transformation layer 160.

[0038] Preferably, vector V2 and vector V6 are all input to mapping layer 170. Mapping layer 170 is used to map vector V2 and vector V6 into the same vector space, and then compare the similarity between them. Mapping layer 170 may essentially be a fully connected layer with a hidden layer, which can map the input vectors into the same vector space according to formula (2) below.

[0039] Y=sigmoid(WX+b) (2) Where, sigmoid represents the activation function sigmoid=1 / (1+exp(-x)), and W and b represent the weights waiting for training. X represents the input vector V2 or V6, with dimensions [batch_size, bert_dim]. Y represents the vector after mapping, with dimensions [batch_size, trans_dim], where trans_dim is the dimension after the entity vector and concept vector are mapped into the same vector space.

[0040] Then, the cosine similarity can be calculated as the semantic similarity based on the following formula (3):

number

[0041] Here, x1 and x2 respectively represent the vector representation of the entity and the vector representation of the concept. In particular, when the mapping layer 170 is used, x1 and x2 respectively represent the vector representations after mapping.

[0042] So far, we have obtained the character similarity and semantic similarity as auxiliary features, which are then input together with the output vector V2 of the BERT model 120 into the splicing layer 130 for subsequent processing, as described above for the main model.

[0043] The training process of the model of the present invention will be described below. For training, a training dataset is input to the model shown in Figure 1, and the training dataset includes text and a knowledge graph. The text includes entities and their contexts, and the knowledge graph includes multiple predefined concepts. The model finally generates classification prediction results for the entities, and the prediction results indicate the concepts to which the entities belong.

[0044] Based on the concept of prediction and the true value, the model is trained by a loss function (e.g., a cross-entropy loss function) to obtain optimal parameters of the model. During training, the initial parameters of the BERT model 120 can be set as the parameters of the BERT pre-trained model. However, those skilled in the art may adopt any other known method to perform training, and the present invention is not limited thereto.

[0045] After training is complete, the trained model can be used to predict specific concepts corresponding to entities mentioned in the text. Hints can then be generated for the user based on the content of the predicted concepts, helping the user understand the exact meaning of the entities. Hints can be provided in various ways (e.g., visually or audibly). For example, while a user is browsing a document, the meaning of an entity can be presented to the user in the form of a hyperlink or a pop-up window, or it can be presented to the user by voice.

[0046] FIG. 4 is a diagram showing an apparatus for presenting hint information to a user in the present invention, and FIG. 5 is a flowchart of a method for presenting hint information to a user in the present invention.

[0047] 4 and 5, the apparatus according to the present invention includes a main module 410, an auxiliary feature generation module 420, and a presentation module 430. In step S510, a combination of an entity, a context, an entity type, and a part of speech of the context (as shown in FIG. 2) is input to the main module 410, and a plurality of predefined concepts are input to the auxiliary feature generation module 420.

[0048] The main module 410 includes a BERT module 411 and a classification module 412. In step S520, the BERT module 411 generates a vector V2 based on a combination of input entities, contexts, entity types, and context parts of speech.

[0049] The auxiliary feature generation module 420 includes a semantic similarity calculation module 421 and a character similarity calculation module 422, and the semantic similarity calculation module 421 further includes a BERT module 4211 and a graph convolutional neural network (GCN) module 4212. Note that the BERT module 4211 in the auxiliary feature generation module 420 and the BERT module 411 in the main module 410 are the same module and may have the same functions, but for convenience, they will be represented by different symbols hereinafter.

[0050] In step S530, the BERT module 4211 generates a vector V4 based on each input concept. In step S540, a graph is generated based on the input concepts and the interrelationships between the concepts, and the GCN module 4212 generates a vector V5 based on the generated graph. The vector V5 indicates the overall features of the concepts, and the overall features represent the interrelationships between the concepts.

[0051] Then, the semantic similarity calculation module 421 concatenates the generated vectors V4 and V5 to generate vector V6 (step S550), and then calculates the semantic similarity between the entity and each concept based on vector V2 and vector V6 using formula (3) (step S560). Preferably, the semantic similarity is calculated after mapping vector V2 and vector V6 to the same vector space.

[0052] In step S570, the character similarity calculation module 422 calculates the character similarity between the entity and each concept.

[0053] In step S580, the classification module 412 generates a classification prediction result based on the vector V2 generated by the BERT module 411 and the semantic similarity and character similarity calculated by the auxiliary feature generation module 420. The prediction result represents the probability that the entity belongs to each concept, and the concept corresponding to the maximum probability can be determined to be the concept to which the entity belongs.

[0054] In step S590, the presentation module 430 generates hint information based on the determined concept to which the entity belongs and presents it to the user. Preferably, the presentation module 430 can present the hint information in a visual and / or auditory manner.

[0055] The method according to the present invention does not necessarily have to be performed in the order shown in Fig. 5, and may be performed in a different order as long as it is technically feasible. For example, step S570 of calculating character similarity may be performed before the process of calculating semantic similarity.

[0056] The model and method of the present invention have been described in detail above with reference to the examples. The present invention provides a new entity disambiguation scheme based on a classification model, in which the BERT model is used, and character similarity and semantic similarity are used as auxiliary features. Furthermore, the present invention uses entity type and context part-of-speech information to enhance semantics, effectively utilizing dictionary resource information and favoring model performance improvement. Furthermore, the present invention uses a graph convolutional neural network (GCN) to extract global features from a graph constructed based on all concepts, thereby integrating information contained in other concepts in the knowledge graph into the current concept, thereby making the semantic information contained in the current concept more complete.

[0057] The methods described in the above embodiments can be realized by software, hardware, or a combination of software and hardware. The programs included in the software can be stored in advance in a storage medium provided inside or outside the device. For example, when executed, these programs can be loaded into a random access memory (RAM) and executed by a processor (e.g., a CPU) to realize the various methods and processes described herein.

[0058] The object of the present invention may also be realized in the following manner: a storage medium storing executable program code is provided directly or indirectly to a related system or device, and a computer or central processing unit (CPU) in the system or device reads and executes the program code. In this case, as long as the system or device has the function of executing a program, the embodiment of the present invention is not limited to a program, and the program may be in any form, such as an object-oriented program, an interpreted program, a script program provided to an operating system, etc.

[0059] The machine-readable storage medium mentioned above may be various types of storage devices and units, semiconductor devices, magnetic disks, optical disks, other media suitable for storing information, and the like.

[0060] Alternatively, the technical solution of the present invention can be realized by connecting a computer to a relevant website on the Internet, downloading and installing the computer program code of the present invention into the computer, and then running the program.

[0061] Each component module and unit in the above-described device may be configured as software, firmware, hardware, or a combination thereof. The specific means and methods used for configuration are well known to those skilled in the art, and therefore will not be described in detail here. When implemented as software or firmware, the programs constituting the software are installed from a storage medium or a network onto a computer having a dedicated hardware structure (e.g., a general-purpose computer 600 shown in FIG. 6), and the computer can realize various functions when various programs are installed.

[0062] FIG. 6 is a structural diagram of a hardware configuration (general-purpose computer) 600 capable of implementing the method and apparatus according to an embodiment of the present invention.

[0063] The general-purpose computer 600 may be, for example, a computer system. Note that the general-purpose computer 600 is merely an example and does not limit the scope or functionality of the method and apparatus according to the present invention. Furthermore, the general-purpose computer 600 does not depend on any module, assembly, or combination thereof in the above-described method and apparatus.

[0064] 6, a central processing unit (CPU) 601 performs various processes based on programs stored in a ROM 602 or programs loaded from a storage unit 608 into a RAM 603. The RAM 603 can also store data required by the CPU 601 to perform various processes, depending on the needs. The CPU 601, ROM 602, and RAM 603 are connected to one another via a bus 604. An input / output interface 605 is also connected to the bus 604.

[0065] The input / output interface 605 is further connected to the following components: an input unit 606 including a keyboard, an output unit 607 including a display such as a liquid crystal display (LCD) and a speaker, a storage unit 608 including a hard disk, and a communication unit 609 including a network interface card such as a LAN card or a modem. The communication unit 609 performs communication processing via a network such as the Internet or a LAN. A drive 610 may be connected to the input / output interface 605 as needed. A removable medium 611, such as a semiconductor memory, can be inserted into the drive 610 as needed to install a computer program read from the medium into the storage unit 608.

[0066] The present invention also provides a program product including machine-readable instruction codes, which, when read and executed by a machine, can perform the methods of the above-described embodiments of the present invention. Accordingly, various storage media for carrying such program products, such as magnetic disks (including floppy disks (registered trademark)), optical disks (including CD-ROMs and DVDs), magneto-optical disks (including MDs (registered trademark)), and semiconductor storage devices, are also included in the present invention.

[0067] The storage medium may include, for example, a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory device, etc., but is not limited to these.

[0068] Furthermore, each operation (process) in the above-described method can also be realized in the form of a computer-executable program stored in various machine-readable storage media.

[0069] Furthermore, the above-mentioned embodiments are further disclosed as supplementary notes as follows.

[0070] (Appendix 1) 1. A computer-implemented method for providing hint information to a user browsing an electronic text using a neural network, comprising: The neural network includes a BERT model and a graph convolutional neural network, and the method includes: inputting the electronic text, information about the electronic text, and a plurality of predefined concepts into the neural network, wherein the electronic text includes entities and contexts of the entities, the information about the electronic text includes types of the entities and parts of speech of the contexts, and the concepts are in text format; generating a first vector (V2) based on a combination of the entity, the context, the type of the entity, and the part of speech of the context using the BERT model; generating a second vector (V4) based on each concept of the plurality of concepts using the BERT model; Using the graph convolutional neural network to generate a third vector (V5) based on a graph, wherein the graph is generated based on the plurality of concepts and the relationships between the concepts; concatenating the second vector (V4) and the third vector (V5) to generate a fourth vector (V6); Calculating a semantic similarity between the entity and each concept among the plurality of concepts based on the first vector (V2) and the fourth vector (V6); determining a concept among the plurality of concepts corresponding to the entity based on the first vector (V2) and the semantic similarity; and generating the hint information based on the determined concepts corresponding to the entities and presenting the hint information to the user.

[0071] (Appendix 2) The method of claim 1, further comprising: Calculating a character similarity between the entity and each concept of the plurality of concepts; and determining a concept corresponding to the entity based on the first vector, the semantic similarity, and the character similarity.

[0072] (Appendix 3) 10. The method of claim 2, the semantic similarity is a cosine similarity calculated based on the first vector (V2) and the fourth vector (V6), and the character similarity is a BM25 similarity between the entity and each concept among the plurality of concepts.

[0073] (Appendix 4) The method of claim 1, further comprising: performing a weighted sum on the entity, the vector corresponding to the context, the entity type, and the vector corresponding to the context part of speech to combine the entity, the context, the entity type, and the context part of speech.

[0074] (Appendix 5) 2. The method of claim 1, comprising: The third vector (V5) represents an overall characteristic of the plurality of concepts, and the overall characteristic represents a mutual relationship between the concepts.

[0075] (Appendix 6) 2. The method of claim 1, comprising: and calculating the semantic similarity after mapping the first vector (V2) and the fourth vector (V6) into the same vector space.

[0076] (Appendix 7) The method of claim 2, further comprising: determining a probability of each concept among the plurality of concepts corresponding to the entity based on the first vector, the semantic similarity, and the character similarity; determining a maximum probability among the determined probabilities as the probability corresponding to the entity when the maximum probability is greater than a predetermined threshold; and The method includes not generating the hint information for the entity when all of the determined probabilities are less than the predetermined thresholds.

[0077] (Appendix 8) 2. The method of claim 1, comprising: The method presents the hint information to the user in at least one of a visual and an auditory manner.

[0078] (Appendix 9) An apparatus for presenting hint information to a user browsing an electronic text using a neural network, comprising: The neural network includes a BERT model and a graph convolutional neural network, and the apparatus: a memory device storing a computer program; and one or more processors coupled to the memory; The processor executes the computer program to implement the method as described in the appendix.

[0079] (Appendix 10) An apparatus for presenting hint information to a user viewing an electronic text, comprising: the electronic text includes entities and contexts of the entities, and the apparatus includes a main module, a semantic similarity calculation module, a character similarity calculation module, and a presentation module; The main module comprises: a BERT module configured to generate a first vector (V2) based on a combination of the entity, the context, the type of the entity, and the part of speech of the context; and a classification module for determining one concept from a plurality of predefined concepts corresponding to the entity based on the first vector (V2), the semantic similarity calculated by the semantic similarity calculation module, and the character similarity calculated by the character similarity calculation module, wherein the concept is in a text format; The semantic similarity calculation module a BERT module configured to generate a second vector (V4) based on each concept of the plurality of concepts; and generating a third vector (V5) based on a graph, wherein the graph includes a graph convolutional neural network generated based on the plurality of concepts and the relationships between the concepts; wherein the semantic similarity calculation module generates a fourth vector (V6) by combining the second vector (V4) and the third vector (V5), and calculates the semantic similarity between the entity and each concept among the plurality of concepts based on the first vector (V2) and the fourth vector (V6); Wherein, the character similarity calculation module is configured to calculate a character similarity between the entity and each concept among the plurality of concepts; The presentation module is configured to generate the hint information based on the determined concept of the entity and present it to the user.

[0080] (Appendix 11) A storage medium storing a computer program, A storage medium, the computer program, when executed by a computer, causing the computer to implement the method for presenting hint information to a user described in any one of Supplementary Notes 1-8.

[0081] Although the preferred embodiment of the present invention has been described above, the present invention is not limited to this embodiment, and any modification to the present invention falls within the technical scope of the present invention as long as it does not depart from the spirit of the present invention.

Claims

1. 1. A computer-implemented method for providing hint information to a user viewing an electronic text using a neural network, comprising: The neural network includes a BERT model and a graph convolutional neural network, and the method includes: inputting the electronic text, information about the electronic text, and text for each of a plurality of concepts in a knowledge graph into the neural network, the electronic text including entities and contexts of the entities, and the information about the electronic text including types of the entities and parts of speech of the contexts; Using the BERT model, a first vector (V ２ ) to generate; A second vector (V) is generated based on the text of each of the concepts in the knowledge graph using the BERT model. ４ ) to generate; The graph convolutional neural network is used to generate a third vector (V ５ ), wherein the graph is generated based on the plurality of concepts and relationships between the plurality of concepts; The second vector (V ４ ) and the third vector (V ５ ) and connect them to form the fourth vector (V ６ ) to generate; The first vector (V ２ ) and the fourth vector (V ６ calculating a semantic similarity between the entity and each concept among the plurality of concepts based on the The first vector (V ２ ) and determining a concept among the plurality of concepts corresponding to the entity based on the semantic similarity; and generating the hint information based on the determined concepts corresponding to the entities and presenting the hint information to the user.

2. 10. The method of claim 1 further comprising: Calculating a character similarity between the entity and each concept of the plurality of concepts; and determining a concept corresponding to the entity based on the first vector (V 2 ), the semantic similarity, and the literal similarity.

3. 3. The method of claim 2, The semantic similarity is expressed as the first vector (V ２ ) and the fourth vector (V ６ ), and the character similarity is a BM25 similarity between the entity and each concept of the plurality of concepts.

4. 10. The method of claim 1, The method further includes performing a weighted sum on the entity, the vector corresponding to the context, the entity type, and the vector corresponding to the context part of speech to combine the entity, the context, the entity type, and the context part of speech.

5. 10. The method of claim 1, The third vector (V ５ ) represents an overall characteristic of the plurality of concepts, and the overall characteristic represents a mutual relationship between the plurality of concepts.

6. 10. The method of claim 1, The first vector (V ２ ) and the fourth vector (V ６ ) into the same vector space before calculating the semantic similarity.

7. 3. The method of claim 2, determining a probability of each concept among the plurality of concepts based on the first vector, the semantic similarity, and the character similarity; When the maximum probability among the determined probabilities is greater than a predetermined threshold, determining the concept with the maximum probability as the concept corresponding to the entity; and The method further includes not determining a concept corresponding to the entity and not generating the hint information for the entity when all of the determined probabilities are less than the predetermined thresholds.

8. 10. The method of claim 1, The method presents the hint information to the user in at least one of a visual manner and an auditory manner.

9. A device for presenting hint information to a user viewing an electronic text using a neural network, comprising: The neural network includes a BERT model and a graph convolutional neural network, and the apparatus: a memory storing a computer program; and a processor coupled to the memory; The processor executes the computer program to inputting the electronic text, information about the electronic text, and text for each of a plurality of concepts in a knowledge graph into the neural network, the electronic text including entities and contexts of the entities, and the information about the electronic text including types of the entities and parts of speech of the contexts; Using the BERT model, a first vector (V ２ ) to generate; A second vector (V) is generated based on the text of each of the concepts in the knowledge graph using the BERT model. ４ ) to generate; The graph convolutional neural network is used to generate a third vector (V ５ ), wherein the graph is generated based on the plurality of concepts and relationships between the plurality of concepts; The second vector (V ４ ) and the third vector (V ５ ) and connect them to form the fourth vector (V ６ ) to generate; The first vector (V ２ ) and the fourth vector (V ６ calculating a semantic similarity between the entity and each concept among the plurality of concepts based on the The first vector (V ２ ) and determining a concept among the plurality of concepts corresponding to the entity based on the semantic similarity; and The hint information is generated based on the determined concept corresponding to the entity, and is presented to the user. An apparatus configured to perform the following:

10. A program for causing a computer to execute the method according to any one of claims 1 to 8.

Images