Apparatus and method for determining at least a part of a knowledge graph

Abstract

To provide a device and method for determining at least one part of a knowledge graph.SOLUTION: In a device 100, for a sentence 112 from a text corpus 110, first to third inputs for first to third classifiers 104-106 are determined. The first input includes a numerical representation of at least one part of the sentence, the second input includes a numerical representation of at least one part of a word, and the third input includes a numerical representation of at least one part of the word. The first classifier determines a numerical representation of a first probability which indicates whether the sentence relates to a knowledge graph 102. If the numerical representation of the first probability satisfies a first condition, the second classifier determines a numerical representation of a second probability which defines a first type for the word from the sentence. The third classifier determines a numerical representation of a third probability which defines a second type for an edge for the word. The word is assigned to a node of the knowledge graph of the first type and connected to another node of the knowledge graph by the edge.SELECTED DRAWING: Figure 1

Description

Technical Field

[0001] Prior Art Luan et al. 2018: Yi Luan et al.: Multi-Task Identification of Entities, Relations, and Coreference for Scientific Knowledge Graph Construction. Conference on Empirical Methods in Natural Language Processing, 2018 discloses means for determining a knowledge graph from a text corpus.

Prior Art Documents

Non-Patent Documents

[0002]

Non-Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0003] Disclosure of the Invention A knowledge graph includes nodes and edges. An edge starts at a node and ends at a node. The representation of the knowledge graph includes, for example, triples, and each triple includes an edge, the node that is the starting point of the edge, and the node that is the ending point of the edge. A node or an edge defines a part of the knowledge graph. According to the apparatus and method described in the independent claims, means for automatically determining at least a part of the knowledge graph from a text corpus are obtained.

Means for Solving the Problem

[0004] According to a method for determining at least a part of a knowledge graph, a text corpus is provided, for sentences from the text corpus, a first input for a first classifier is determined, a second input for a second classifier is determined, a third input for a third classifier is determined, the first input includes a numerical representation of at least a part of the sentence, the second input includes a numerical representation of at least a part of a word from the sentence, the third input includes a numerical representation of at least a part of a word from the sentence, by the first classifier, according to the first input, a numerical representation of a first probability indicating whether the sentence is related to the knowledge graph is determined, when the numerical representation of the first probability satisfies a first condition, by the second classifier, according to the second input, a numerical representation of a second probability defining a first type for a word from the sentence is determined, by the third classifier, according to the third input, a numerical representation of a third probability defining a second type for an edge for the word is determined, words from the sentence are associated with nodes of a knowledge graph of the first type and are intended to be connected to other nodes of the knowledge graph by edges of the second type. Thereby, for example, it becomes possible to automatically extract information from free text from publications in the field of materials science, that is, text without manual annotation, and thereby it becomes possible to construct a knowledge graph. In this case, relationships that are often not explicitly described in the text are interpreted. The edge corresponding to this relationship is predicted by the fact that the end node of this edge corresponds to words of a specific first type and a specific second type. The latter second type defines the edge in this example. The start node of the edge corresponds to another word from the sentence in this example, for example, a word associated with a specific first type. The edge is associated with two words at the start and end of the edge.

[0005] Preferably, the first word of the sentence is mapped onto a first numerical representation by a first function, the first word is mapped onto a second numerical representation by a second function different from the first function, the second word of the sentence is mapped onto a third numerical representation by the first function, the second word is mapped onto a fourth numerical representation by the second function, and the first numerical representation, the second numerical representation, the third numerical representation, and the fourth numerical representation are mapped by a third function onto a first tensor defining a first input and / or a second input. Thereby, a plurality of words from the sentence are used to classify the sentence, and at least two different numerical representations, i.e., embeddings, are determined for each word. Thereby, the classification of the sentence is improved.

[0006] Preferably, training data including a plurality of tuples is prepared, and in each tuple, a value of a particularly binary variable defining whether the sentence of the text corpus is related to the knowledge graph is associated with the sentence of the text corpus, a first label defining a first type is associated, a second label defining a second type is associated, and at least one parameter of the first function, the second function, the third function, the first classifier, the second classifier, and / or the third classifier is trained based on the training data. These tuples can include words annotated by an expert from the text corpus.

[0007] Training data including a plurality of tuples can be prepared, and in each tuple, a numerical representation of a first probability regarding the sentence, a numerical representation of a second probability regarding a word from the sentence, and a numerical representation of a third probability regarding a word from the sentence are associated with the sentence of the text corpus, and at least one parameter of the first function, the second function, the third function, the first classifier, the second classifier, and / or the third classifier is trained based on the training data. These tuples can include the positions of words defining words annotated by an expert from the text corpus.

[0008] The first classifier can include a first layer that determines a vector for a sentence in response to a first input. The first classifier includes a second layer that determines a numerical representation of a first probability, particularly in binary, in response to the vector. This enables classification for each sentence.

[0009] The second classifier can include a first layer that determines a vector for a word from a sentence in response to a second input, and the second classifier includes a second layer that determines a plurality of numerical representations of a second probability in response to the vector, where each of the plurality of numerical representations is associated with a type of node for a knowledge graph. This enables tagging of sequences with a specific vocabulary that defines these types.

[0010] Words from sentences in a text corpus can be associated with nodes of a knowledge graph when the word is associated with a type of node of the knowledge graph. This enables creation of a large number of nodes for the knowledge graph.

[0011] The third classifier can include a first layer that determines a vector for a word from a sentence in response to a third input, and the third classifier includes a second layer that determines a plurality of numerical representations of a third probability in response to the vector, where each of the plurality of numerical representations is associated with a type of edge for a knowledge graph. The knowledge graph represents, for example, an experiment by having other nodes representing materials or equipment provided in the experiment associated with the experiment node. The third classifier identifies the type of edge with respect to an edge between the experiment node and one of the other nodes. The type of edge can define the use of materials in the experiment. This enables tagging of sequences with a specific vocabulary that defines these types.

[0012] Preferably, the first input, the second input, and / or the third input for the sentence are determined according to a plurality of words from the sentence. By considering only the words from the same sentence, the amount of input data to be processed is reduced, making it easier to determine the knowledge graph.

[0013] An apparatus for determining at least a part of the knowledge graph is configured to implement the method.

[0014] Further advantageous embodiments will become apparent from the following description and the drawings.

Brief Description of the Drawings

[0015]

Figure 1

Figure 2

Figure 3

Embodiments for Carrying Out the Invention

[0016] FIG. 1 schematically shows an apparatus 100 for determining at least a part of a knowledge graph 102. The knowledge graph 102 is defined by triples that define edges, nodes that are the starting points of the edges, and nodes that are the ending points of the edges. A part of the knowledge graph is a node or an edge.

[0017] Regarding the procedure using the classifier described below, for the following sentence: The SOFC with Pt / SmNiO3 demonstrated dramatic power output. will be described based on.

[0018] The first classifier 104 identifies that this sentence is relevant, i.e., that this sentence contains information to be written into the knowledge graph.

[0019] The second classifier 106 identifies concepts. Concepts can relate to various different types. In this sentence, for example, it is identified that SOFC relates to the type "device (Geraet)", Pt relates to the type "material", SmNiO3 relates to the type "material", and demonstrated relates to the type "verb describing an experiment". Words of the type "verb describing an experiment" are defined as nodes that can be the starting point of an edge in the knowledge graph. Such nodes describing an experiment are hereinafter referred to as experiment nodes.

[0020] In this example, all concepts identified by the second classifier are accommodated as nodes in the knowledge graph.

[0021] The third classifier 108 identifies that SOFC relates to the type "experiment device", Pt relates to the type "anode material of the experiment", and SmNiO3 relates to the type "fuel used in the experiment". These words define nodes that are the end points of edges starting from the experiment node. The identified types define the labels of the edges.

[0022] In the knowledge graph, the SOFC node is connected to the experiment node and written on the edge "device". Similarly, the Pt node is connected to the experiment node and written on the edge "anode material".

[0023] The apparatus 100 includes the first classifier 104, the second classifier 106, and the third classifier 108.

[0024] Device 100 is configured to prepare a text corpus 110. The text corpus includes a plurality of sentences 112. Each sentence 112 includes a plurality of words 112-1, ···, 112-n. The number n of words may be different in each sentence 112. Device 100 is configured to prepare a first input for the first classifier 104, a second input for the second classifier 106, and a third input for the third classifier 108 for the sentences 112 from the text corpus 110.

[0025] The first classifier 104 includes a first layer 104-1 configured to determine a vector for the sentence 112 according to the first input. The first classifier 104 includes a second layer 104-2 configured to determine a numerical representation of a first probability according to the vector. In this example, the classification distinguishes between binaries, that is, between two classes. The probability itself is, in this example, not binary, that is, not 0 or 1, but a numerical probability value.

[0026] Device 100 is configured to map the first word of the sentence 112 onto a first numerical representation by a first function. Device 100 can be configured to map the first word onto a second numerical representation by a second function different from the first function. The first function maps the word according to the parameters of the first function, for example, to the following embeddings: word2vec, mat2vec, bpe, BERT, SciBERT to one of them.

[0027] The second function maps the word according to the parameters of the second function, for example, to the following embeddings: word2vec, mat2vec, bpe, BERT, SciBERT to another one different from the above.

[0028] In this example, an embedding layer is arranged in the device 100 for this purpose, and the words of sentence 112 can be mapped using this embedding layer.

[0029] In this example, a plurality of pre-trained types of embedding layers, namely, word2vec, mat2vec, byte-pair-encoding embedding, BERT embedding, and SciBERT embedding, are combined.

[0030] Word2vec is described, for example, in Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. 2013. Efficient estimation of word representations in vector space. In ICLR Workshop.

[0031] Mat2vec is described, for example, in Vahe Tshitoyan, John Dagdelen, Leigh Weston, Alexander Dunn, Ziqin Rong, Olga Kononova, Kristin A. Persson, Gerbrand Ceder, and Anubhav Jain. 2019. Unsupervised word embeddings capture latent knowledge from materials science literature. Nature, 571:95 - 98.

[0032] Byte-pair encoding is described, for example, in Benjamin Heinzerling and Michael Strube. 2018. BPEmb: Tokenization-free Pre-trained Subword Embeddings in 275 Languages. In Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018), Miyazaki, Japan. European Language Resources Association (ELRA).

[0033] BERT embeddings are described, for example, in Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4171-4186, Minneapolis, Minnesota. Association for Computational Linguistics.

[0034] The SciBERT embedding is described, for example, in Iz Beltagy, Kyle Lo, and Arman Cohan. 2019. Scibert: A pretrained language model for scientific text. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 3606-3611.

[0035] Since the last three are subword-based embeddings, they are particularly well-suited for representing complex long words. Mat2vec is an embedding trained on materials science text using the word2vec algorithm, so it is particularly well-suited for representing domain-specific words.

[0036] The apparatus 100 is configured to map a second word onto a third numerical representation by a first function. The second word is another word of the same sentence 112.

[0037] The apparatus 100 can be configured to map the second word onto a fourth numerical representation by a second function.

[0038] The first numerical representation and the third numerical representation can be mapped by a third function onto a first tensor defining a first input and / or a second input. The first numerical representation, the second numerical representation, the third numerical representation, and the fourth numerical representation can be mapped by a third function onto a first tensor defining a first input and / or a second input.

[0039] The third function can include a concatenation of numerical representations, and by this concatenation, a first tensor is formed.

[0040] The second classifier 106 includes a first layer 106-1 configured to determine a vector for a word from sentence 112 in response to a second input. The second classifier 106 includes a second layer 106-2 configured to determine a plurality of numerical representations of a second probability in response to the vector.

[0041] The third classifier 108 includes a first layer 108-1 configured to determine a vector for a word from sentence 112 in response to a third input. The third classifier 108 includes a second layer 108-2 configured to determine a plurality of numerical representations of a third probability in response to the vector.

[0042] The apparatus 100 is configured to associate a word from sentence 112 with a node of the knowledge graph 102 when the word is associated with a first type.

[0043] The apparatus 100 is configured to associate a word from sentence 112 with a node that is the end point of an edge of the knowledge graph 102 when the word is associated with a second type.

[0044] The apparatus 100 is configured to classify sentence 112 from the text corpus 110 as to whether the sentence 112 is relevant for the determination of the knowledge graph.

[0045] For this purpose, the apparatus 100 is configured to implement the method described below for determining at least a part of the knowledge graph.

[0046] The apparatus 100 can include at least one processor and at least one memory configured to implement the method described below for determining at least a part of the knowledge graph.

[0047] The first classifier 104 can be implemented as a BiLSTM model, i.e., as a bidirectional long short-term memory.

[0048] The second classifier 106 can be implemented as a BiLSTM-CRF model, that is, as a bidirectional long short-term memory with a conditional random field layer.

[0049] The third classifier 108 can be implemented as a BiLSTM-CRF model, that is, as a bidirectional long short-term memory with a conditional random field layer.

[0050] In this example, the method for determining at least a part of the knowledge graph 102 is carried out after the training of the classifier. It is also possible to carry out this method independently of the training step after training.

[0051] In step 200, training data is prepared. The training data includes a plurality of tuples in this example.

[0052] In one aspect, in each tuple, for the sentence 112 of the text corpus 110, a value of a particularly binary variable that defines whether the sentence is related to the knowledge graph is associated, a first label that defines the first type is associated, and a second label that defines the second type is associated.

[0053] In other aspects, in each tuple, for the sentence 112 of the text corpus 110, a numerical representation of a first probability regarding the sentence 112, a numerical representation of a second probability regarding the first type, and a numerical representation of a third probability regarding the second type are associated.

[0054] During training, at least one parameter of the first function, the second function, the third function, the first classifier 104, the second classifier 106, and / or the third classifier 108 is trained based on the training data.

[0055] In this example, the parameters of the first function, the second function, the third function, the first classifier 104, the second classifier 106, and the third classifier 108 are determined using the gradient descent method such that the sum of the first deviation related to the first classifier, the second deviation related to the second classifier, and the third deviation related to the third classifier satisfies a predetermined condition. In this example, the condition is satisfied when a certain sum value is smaller than or is the minimum value of other sum values determined using other parameters.

[0056] In this example, the first deviation is defined by a first difference between a first numerical representation determined for a word from a tuple and a value related thereto from the tuple.

[0057] In this example, the second deviation is defined by a second difference between a second numerical representation determined for a word from a tuple and a value related thereto from the tuple.

[0058] In this example, the third deviation is defined by a third difference between a third numerical representation determined for a word from a tuple and a value related thereto from the tuple.

[0059] The training data is separable into a plurality of parts. In an iteration, step 200 can be repeated for each individual part of the training data.

[0060] In step 202, the text corpus 110 is prepared.

[0061] In step 204, a first input for the first classifier 104 is determined for the sentence 112 from the text corpus 110. The first input includes a numerical representation of at least a part of the sentence 112.

[0062] In this example, in step 204-11, the first word of sentence 112 is mapped onto a first numerical representation by a first function. In step 204-21, the first word can be mapped onto a second numerical representation by a second function.

[0063] In this example, in step 204-12, the second word of sentence 112 is mapped onto a third numerical representation by a first function. In step 204-22, the second word can be mapped onto a fourth numerical representation by a second function.

[0064] In step 204-3, the first numerical representation, the second numerical representation, the third numerical representation, and the fourth numerical representation can be mapped by a third function onto a first tensor that defines a first input. The first tensor can also be determined according to all n words of the sentence or a plurality of words of the sentence. m functions can be provided, and by the m functions, m different numerical representations for at least a part of one word are determined. This is schematically shown in FIG. 3.

[0065] In step 206, a second input for the second classifier 106 is determined. The second input includes numerical representations of at least a part of the words from sentence 112.

[0066] In this example, the first tensor defines a second input for the second classifier 106.

[0067] In step 208, a third input for the third classifier 108 is determined.

[0068] The third input is the same as the second input in this example.

[0069] In step 210, by the first classifier 104, a numerical representation of a first probability indicating whether sentence 112 is related to the knowledge graph 102 is determined according to the first input.

[0070] In this example, the value of the numerical representation of the first probability is determined between 0 and 1.

[0071] In this example, the first classifier 104 determines, by the first layer 104-1, a vector for sentence 112 according to the first input, and determines, by the second layer 104-2, the numerical representation of the first probability according to the vector.

[0072] In this sentence classification, the attention layer of the artificial neural network having a BiLSTM structure can calculate the sentence representation as a weighted representation of the hidden state of the BiLSTM, and can classify the softmax layer in the binary classification as to whether the sentence describes information relevant to the knowledge graph.

[0073] In step 212, it is checked whether the numerical representation of the first probability satisfies the first condition. In this example, the condition is satisfied when the value of the numerical representation exceeds the threshold. This means that, that is, when it is determined that the sentence is related to the knowledge graph.

[0074] When the numerical representation of the first probability satisfies the first condition, step 214 is executed. Otherwise, step 204 is executed for other sentences from the text corpus 110.

[0075] In step 214, the second classifier 106 determines, according to the second input, the numerical representation of the second probability indicating the first type.

[0076] The second classifier 106 determines, by the first layer 106-1, a vector for the words from sentence 112 according to the second input. In this example, the second classifier 106 determines, by the second layer 106-2, a plurality of numerical representations of the second probability according to the vector, and each numerical representation among the plurality of numerical representations is associated with the type of the node for the knowledge graph 102.

[0077] Examples of node types are MATERIAL, VALUE, and DEVICE in the case of a knowledge graph that describes one or more materials science experiments.

[0078] Words of the type "MATERIAL" include, for example, chemical formulas or names of compounds such as oxides or hydrocarbons.

[0079] Words of the type "VALUE" include, for example, measures such as 750 °C or comparisons such as greater than, equal to, between.

[0080] Words of the type "DEVICE" include, for example, names of devices such as machines, instruments, or domain-specific abbreviations for devices.

[0081] The third classifier 108 determines a numerical representation of a third probability indicating a second type in response to a third input.

[0082] The third classifier 108 determines, by means of a first layer 108-1, a vector for a word from sentence 112 in response to the third input. The third classifier 108 determines, by means of a second layer 108-2, a plurality of numerical representations of a third probability in response to the vector. Each of the numerical representations among the plurality of numerical representations corresponds, in this example, to an edge type for the knowledge graph 102.

[0083] Examples of edge types are anode material, cathode material, fuel in the case of this knowledge graph.

[0084] Words of the type "anode material" include, for example, word components indicating materials that can be used as anodes.

[0085] Words of the type "fuel" include, for example, word components indicating components of the fuel.

[0086] In step 216, the first word from sentence 112 is associated with a node in the knowledge graph 102. It is also possible to associate the second word from sentence 112 with the node. The first word indicates, for example, the material described in the sentence. The second word indicates, for example, the equipment used in the experiment.

[0087] In optional step 218, it is possible to check whether the node is an experiment node. If the node is an experiment node, step 204 is performed on other words from the sentence. Otherwise, step 220 is performed.

[0088] In the knowledge graph, in this example in step 220, an edge is inserted between the node and the experiment node. The edge can be labeled with a label defined by the identified type of the edge. Then, step 204 is performed on other sentences from the text corpus 110.

[0089] In this example, the words from sentence 112 in the text corpus 110 are associated with a node in the knowledge graph 102 when the word is associated with a first type to which it corresponds. In this case, this node is connected to the experiment node by an edge of a second type.

[0090] In this example, the method ends when all sentences from the text corpus 110 have been processed. The method can also be performed on only a predefined number of sentences from the text corpus 110.

[0091] This enables, for example, the automatic extraction of information from publications in the field of materials science that are not annotated, thereby enabling the construction of a knowledge graph. For this purpose, it is contemplated to automatically identify relevant sentences from a text corpus. A sentence is relevant if it describes, for example, a materials science experiment. Further, it is contemplated to automatically identify words having relevance as nodes in the knowledge graph from a sentence. A word has relevance to a node if, for example, the word describes a concept particularly related to a materials science experiment. Examples of concepts from the field of materials science are materials or physical units. Further, it is contemplated to automatically identify from the text corpus words to be connected to an experiment node by an edge in the knowledge graph. The edge is entered in the knowledge graph, for example, with a label indicating what role a certain concept plays particularly in a materials science experiment. Examples of roles from the field of materials science are anode material, cathode material. This approach is applicable to other domains other than materials science. The classifier can be configured as one or more models, particularly as an artificial neural network, trained to be able to automatically extract information regarding experiments from publications in a scientific field and write it into the knowledge graph. The first classifier can be configured as a binary classification model, particularly trained to indicate binary whether a sentence is relevant to an experiment. The second classifier and the third classifier can each be implemented as a sequence tagging model. The sequence tagging model classifies words regarding which of a plurality of pre-defined labels should be associated with the word having the highest probability among these labels. The labels for the second classifier can be of a type defining possible concepts. The labels for the third classifier can be of a type defining possible roles. The possible roles and possible concepts can be words from a word set defined for the domain to which the text corpus is associated.For example, by repeating the steps for all words from a text corpus, concepts having relevance as nodes, i.e., a first set of words from the text corpus, are determined, and roles having relevance as nodes that are endpoints of edges, i.e., a second set of words from the text corpus, are determined. Words for the nodes thus discovered can be included in the knowledge graph and connected to the experimental nodes by edges. Thus, the knowledge graph can be used by domain experts to search for information and related work regarding a particular experiment as intended.

Claims

1. A method for determining at least a part of a knowledge graph (102), comprising: A text corpus (110) is provided (202); For sentences (112) from the text corpus (110), A first input for a first classifier (104) is determined (204); A second input for a second classifier (106) is determined (206); A third input for a third classifier (108) is determined (208); The first input includes a numerical representation of at least a part of the sentence (112), the second input includes a numerical representation of at least a part of a word from the sentence (112), and the third input includes a numerical representation of at least a part of the word from the sentence (112); Based on the first input, the first classifier (104) determines a numerical representation of a first probability indicating whether the sentence (112) is related to the knowledge graph (102) (210); When the numerical representation of the first probability satisfies a first condition (212), based on the second input, the second classifier (106) determines a numerical representation of a second probability defining a first type for the word from the sentence (112) (214); Based on the third input, the third classifier (108) determines a numerical representation of a third probability defining a second type for an edge for the word (214); The word from the sentence (112) is associated with a node of the knowledge graph (102) of the first type and is connected to other nodes of the knowledge graph by an edge of the second type (220); A method.

2. The first word of the sentence (112) is mapped onto a first numerical representation by a first function (204-11); The first word is mapped onto a second numerical representation by a second function different from the first function (204-21); The second word of the sentence (112) is mapped onto a third numerical representation by the first function (204-12); The second word is mapped onto a fourth numerical representation by the second function (204-22); The first numerical representation, the second numerical representation, the third numerical representation, and the fourth numerical representation are mapped onto a first tensor defining the first input and / or the second input by a third function (204-3); The method according to claim 1.

3. Training data including a plurality of tuples is prepared, In each tuple, for the sentence (112) of the text corpus (110), a value of a particularly binary variable that defines whether the sentence is related to the knowledge graph is associated, a first label that defines the first type is associated, and a second label that defines the second type is associated, At least one parameter of the first function, the second function, the third function, the first classifier (104), the second classifier (106), and / or the third classifier (108) is trained (200) based on the training data, The method according to claim 2.

4. Training data including a plurality of tuples is prepared, In each tuple, for the sentence (112) of the text corpus (110), a numerical representation of the first probability regarding the sentence (112), a numerical representation of the second probability regarding the word from the sentence, and a numerical representation of the third probability regarding the word from the sentence (112) are associated, At least one parameter of the first function, the second function, the third function, the first classifier (104), the second classifier (106), and / or the third classifier (108) is trained (200) based on the training data, The method according to claim 2 or 3.

5. The first classifier (104) includes a first layer (104-1) that determines a vector for the sentence (112) according to the first input, The first classifier (104) includes a second layer (104-2) that determines a numerical representation of the first probability according to the vector, The method according to any one of claims 1 to 4.

6. The second classifier (106) includes a first layer (106-1) that determines (214-1) a vector for the word from the sentence (112) according to the second input, The second classifier (106) includes a second layer (106-2) that determines (214-2) a plurality of numerical representations of the second probability according to the vector, Each of the plurality of numerical representations is associated with a type of node for the knowledge graph (102), The method according to any one of claims 1 to 5.

7. Words from the sentence (112) from the text corpus (110) are associated (218) with a node of the knowledge graph (102) when the word is associated with the type of a node of the knowledge graph (102). The method according to claim 6.

8. The third classifier (108) includes a first layer (108-1) that determines a vector for the word from the sentence (112) according to the third input. The third classifier (108) includes a second layer (108-2) that determines a plurality of numerical representations of the third probability according to the vector. Each of the plurality of numerical representations is associated with the type of an edge for the knowledge graph (102). The method according to any one of claims 1 to 7.

9. The first input, the second input, and / or the third input for the sentence (112) are determined (208) according to a plurality of words from the sentence (112). The method according to any one of claims 1 to 8.

10. An apparatus (100) for determining at least a part of a knowledge graph (102), An apparatus (100) characterized by being configured to implement the method according to any one of claims 1 to 9.

11. A computer program, The computer program includes computer-readable instructions for implementing the method according to any one of claims 1 to 9 when executed by a computer.

Images