Information processing device, information processing method, and program

Abstract

We provide link prediction technology that enables lower-cost and more accurate predictions. [Solution] The information processing device includes an acquisition unit that acquires a query, a first generation unit that generates one or more triples by referring to the query, a second generation unit that generates a text from each of the one or more triples, a first calculation unit that acquires text embedding obtained using a trained natural language model for each of the generated texts and calculates a score for each of the text embeddings, a second calculation unit that calculates a score for the knowledge graph embedding of each of the one or more triples, and an aggregation unit that aggregates the scores calculated by the first calculation unit and the scores calculated by the second calculation unit.

Description

Technical Field

[0006] , ,

[0005] , , ,

[0007] , , ,

[0001] The present invention relates to an information processing apparatus, an information processing method, and a program.

Background Art

[0002] Various link prediction techniques using a knowledge graph for predicting genes (proteins) related to diseases such as cancer have been provided.

[0003] For example, Patent Document 1 discloses a technique for realizing link prediction after learning a natural language model from automatically extracted triples. Patent Document 2 also discloses a technique for learning nodes and documents in a knowledge graph.

Prior Art Documents

Non-Patent Documents

[0004]

Patent Document 1

Patent Document 2

Summary of the Invention

Problems to be Solved by the Invention

[0005] However, when using such a learned model, it is necessary to prepare training data with a sufficient amount of data such as sentences and documents, or a natural language model with a high learning cost, resulting in a problem of high learning cost.

[0006] [[ID=4​​​​​​​An information processing device relating to one aspect of this disclosure includes: an acquisition means for acquiring a query; a first generation means for generating one or more triples by referring to the query; a second generation means for generating a text from each of the one or more triples; a first calculation means for acquiring text embeddings obtained using a trained natural language model for each of the generated texts and calculating a score for each of the text embeddings; a second calculation means for calculating a score for the knowledge graph embedding of each of the one or more triples; and an aggregation means for aggregating the scores calculated by the first calculation means and the scores calculated by the second calculation means.

[0008] An information processing device relating to one aspect of this disclosure includes: a first generation means for generating a group of triples including positive example triples and negative example triples by referring to a knowledge graph; a second generation means for generating a sentence from each of the triples included in the group of triples; and a first learning means for training a score calculation model that obtains sentence embeddings obtained using a trained natural language model for each of the generated sentences and calculates a score for each of the sentence embeddings by referring to each of the sentence embeddings. It is equipped with.

[0009] An information processing method relating to one aspect of this disclosure includes obtaining a query, generating one or more triples by referring to the query, generating a text from each of the one or more triples, obtaining a text embedding for each of the generated texts using a trained natural language model and calculating a score for each of the text embeddings, calculating a score for the knowledge graph embedding for each of the one or more triples, and aggregating the scores of the text embeddings and the knowledge graph embeddings.

[0010] An information processing method relating to one aspect of this disclosure includes: generating a group of triples including positive example triples and negative example triples by referring to a knowledge graph; generating a sentence from each of the triples included in the group of triples; and for each of the generated sentences, obtaining a sentence embedding obtained using a trained natural language model, and training a scoring model that calculates a score for each of the sentence embeddings by referring to each of the sentence embeddings.

[0011] A program relating to one aspect of this disclosure is a program for causing a computer to function as an information processing device, and causes the computer to function as the first generation means, the second generation means, the first calculation means, the second calculation means, and the aggregation means.

[0012] A program relating to one aspect of this disclosure is a program for causing a computer to function as an information processing device, and causes the computer to function as the first generation means, the second generation means, and the first learning means. [Effects of the Invention]

[0013] This disclosure provides a link prediction technology that enables lower-cost and more accurate predictions. [Brief explanation of the drawing]

[0014] [Figure 1] This is a block diagram showing the configuration of the information processing device related to this disclosure. [Figure 2] This is a flowchart showing the flow of the information processing method related to this disclosure. [Figure 3] This diagram illustrates the basic structure of a knowledge graph. [Figure 4] This is a block diagram showing the configuration of the information processing device related to this disclosure. [Figure 5] This is a flowchart showing the flow of the information processing method related to this disclosure. [Figure 6] This is a block diagram showing the configuration of the information processing system related to this disclosure. [Figure 7] It is a flowchart showing the flow of the information processing method according to the present disclosure. [Figure 8] It is a flowchart showing the flow of the information processing method according to the present disclosure. [Figure 9] It is a diagram for explaining the information processing according to the present disclosure. [Figure 10] It is a diagram for explaining the information processing according to the present disclosure. [Figure 11] It is a diagram for explaining the information processing according to the present disclosure. [Figure 12] It is a block diagram showing the configuration of the information processing apparatus according to the present disclosure. [Figure 13] It is a block diagram showing the hardware configuration of the information processing apparatus according to the present disclosure.

Embodiments for Carrying Out the Invention

[0015] Hereinafter, embodiments of the present invention will be exemplified. However, the present invention is not limited to the following exemplary embodiments, and various modifications are possible within the scope shown in the claims. For example, embodiments obtained by appropriately combining the technologies (part or all of the things or methods) employed in the following exemplary embodiments may also be included in the scope of the present invention. Also, embodiments obtained by appropriately omitting a part of the technologies employed in the following exemplary embodiments may also be included in the scope of the present invention. Further, the effects mentioned in the following exemplary embodiments are merely examples of the effects expected in those exemplary embodiments and do not define the extension of the present invention. That is, embodiments that do not exhibit the effects mentioned in the following exemplary embodiments may also be included in the scope of the present invention.

[0016] 〔First Embodiment〕 A first exemplary embodiment, which is an example of an embodiment of the present invention, will be described in detail with reference to the drawings. This exemplary embodiment is the basic form for each of the exemplary embodiments described later. The scope of application of each technology adopted in this exemplary embodiment is not limited to this exemplary embodiment. That is, each technology adopted in this exemplary embodiment can also be adopted in other exemplary embodiments included in this disclosure, to the extent that no particular technical problems occur. Furthermore, each technology shown in the drawings referenced to explain this exemplary embodiment can also be adopted in other exemplary embodiments included in this disclosure, to the extent that no particular technical problems occur.

[0017] (Configuration of Information Processing Device 1) The configuration of the information processing device 1 according to this exemplary embodiment will be described with reference to Figure 1. Figure 2 is a block diagram showing the configuration of the information processing device 1. As shown in Figure 1, the information processing device 1 includes an acquisition unit (means) 11, a first generation unit (means) 12, a second generation unit 13 (means), a first calculation unit (means) 14, a second calculation unit (means) 15, and an aggregation unit (means) 16.

[0018] (Acquisition part 11) The acquisition unit 11 acquires queries entered by the user. Here, the query may include, for example, multiple information fragments, each representing one or more targets, and each representing a relationship with any of those targets. The acquisition unit 11 extracts these information fragments from the acquired query. For example, if the acquisition unit 11 acquires the query "Which proteins are highly associated with disease 1?", it extracts "disease 1" as an information fragment representing a target and "highly associated" as an information fragment representing a relationship with that target. The acquisition unit may also further extract "protein" as an information fragment representing a target.

[0019] The queries retrieved by the retrieval unit 11 may be expressed using knowledge graph terminology as follows: An example of such a query retrieved by the retrieval unit 11 is a sentence where one of the components of a triple is missing, or a sentence where one of the components of a triple is relatively abstract. For example, in the above example, "protein" is relatively abstract, and there are various concrete examples of protein; therefore, these concrete examples can be the answer to the query.

[0020] Here, the triple described above is defined by a link as a component of the knowledge graph and the nodes at both ends of that link. For example, if the knowledge graph is configured as a directed graph, the triple is composed of (a node as the head, a link as the relation, and a node as the tail).

[0021] The following explains knowledge graphs and triples with reference to Figure 3. Figure 3 shows an example of a knowledge graph. A knowledge graph is a type of data structure, and in the example shown in Figure 3, it is composed of nodes (nodes 1-5) and links (links 1-5). A knowledge graph can also be described as a data format that represents the connections between people and things. Nodes are sometimes called vertices, vertices, or entities, and links are sometimes called relationships, edges, or relations.

[0022] The smallest record unit that makes up a knowledge graph is a triplicate (hereinafter referred to as a triple) representing the constituent elements and their relationships, and a knowledge graph is constructed by enumerating these triples. The constituent elements at both ends (head, tail) correspond to the nodes mentioned above, and the relationship in the middle corresponds to the links mentioned above. The order within the triples in Figure 3 is indicated by the direction of the arrows.

[0023] An example of a query retrieved by the retrieval unit 11 is as shown in the example above: In addition to descriptions such as "Which proteins are highly associated with disease 1?" for node prediction, • This may include descriptions such as "What is the relationship between disease 2 and protein 1?" to perform link prediction.

[0024] The format of the above query is not particularly limited; it may be written in a format that can be applied to any given format, such as listing only the head and tail of a triple, or only the head and relation, or the format may be changed as appropriate.

[0025] (First generation unit 12) The first generation unit 12 generates one or more triples by referring to the query. Here, one or more triples that can be expected as prediction results are generated based on the query entered by the user. For example, if the query is "Which protein is highly associated with disease 1?", the first generation unit 12 generates one or more triples by replacing missing components or relatively abstract components with specific examples (protein 1, protein 2, ...), such as (disease 1, highly associated, protein 1), (disease 1, highly associated, protein 2), ... The number of triples that the first generation unit 12 generates for a given triple may be a predetermined number (e.g., 100), or it may be changed according to the acquired triple.

[0026] (Second generation unit 13) The second generation unit 13 generates sentences from each of the one or more triples. For example, for the triple (disease 1, highly related, protein 1), it generates a sentence such as "Disease 1 and protein 1 are highly related," and for the triples (drug 1, prescribe, disease 1) and (drug 2, prescribe, disease 1), it generates sentences such as "prescribe drug 1 for disease 1" and "prescribe drug 2 for disease 1," respectively.

[0027] (First calculation unit 14) The first calculation unit 14 obtains a sentence embedding for each generated sentence using a trained natural language model and calculates a score for each sentence embedding. Here, "obtaining a sentence embedding" refers, for example, to obtaining a vector corresponding to the sentence (vectorizing the sentence). Specific examples of trained natural language models are not limited to this exemplary embodiment, but examples include Sentence-BERT and LLM. The natural language model may, for example, acquire domain knowledge by training primarily on literature in the target field, such as medicine or biomedicine. By performing sentence embedding using such a natural language model, the data of the generated sentences is vectorized, and their semantic closeness is quantified. Based on this quantified value, a score for each sentence embedding can be calculated.

[0028] (Second calculation unit 15) The second calculation unit 15 calculates the knowledge graph embedding score for each of the one or more triples. Knowledge graph embedding is a method of embedding the components (triples) of the knowledge graph as vectors, and the score is calculated based on the vector values ​​obtained here. In other words, "calculating the knowledge graph embedding score of a triple" means, for example, vectorizing the triple and calculating the score of the vector. Note that the knowledge graph embedding process can be performed using a trained embedding model as an example.

[0029] (Aggregation section 16) The aggregation unit 16 aggregates the scores calculated by the first calculation unit and the scores calculated by the second calculation unit. This aggregation allows for highly accurate scoring (i.e., prediction) of the validity of the existence of each triple. The prediction results from the aggregation unit 16 are, for example, visually presented to the user via a display unit (not shown) or provided to other devices via a communication unit (not shown). For score aggregation, for example, the average of the two scores may be taken, or a weighted average may be taken.

[0030] (Effects of Information Processing Device 1) As described above, in the information processing device 1, • Retrieve the query, • Generate one or more triples by referring to the above query, • Generate a sentence from each of the one or more triples mentioned above. For each generated sentence, obtain a sentence embedding using a trained natural language model, and calculate the score for each of those sentence embeddings. • Calculate the score for each of the one or more triples of the knowledge graph embedding. - The score calculated by the first calculation means and the score calculated by the second calculation means are aggregated. This configuration is adopted.

[0031] Thus, in the information processing device 1, • Generate a sentence from each of the one or more triples mentioned above. For each generated sentence, obtain a sentence embedding using a trained natural language model, and calculate a score for each of those sentence embeddings. Because this configuration is adopted, it is possible to suppress increases in costs (time costs and expenses). For example, it is possible to suppress increases in costs compared to a configuration that generates a large number of sentences using a pre-trained natural language model.

[0032] Furthermore, in the information processing device 1, - The score calculated by the first calculation means and the score calculated by the second calculation means are aggregated. Because this configuration is employed, high-precision predictions are possible by utilizing two vector information sources: text embedding and knowledge graph embedding. Therefore, this configuration provides a link (or node) prediction technology that enables lower-cost and higher-precision predictions.

[0033] (Information processing method S1 flow) Next, the flow of the information processing method S1 according to this exemplary embodiment will be explained with reference to Figure 2. Figure 2 is a flowchart showing the flow of the information processing method S1. As shown in Figure 2, the information processing method S1 includes a step (processing) S11 for obtaining a query, a step (processing) S12 for generating triples, a step (processing) S13 for generating text from triples, a step (processing) S14 for obtaining text embedding and calculating the score of the text embedding, a step (processing) S15 for calculating the score of knowledge graph embedding, and a step (processing) S16 for aggregating the scores.

[0034] (Step S11) In step S11, the acquisition unit 11 retrieves the query. A more detailed explanation of the acquisition unit 11 has been given above, so it will be omitted here.

[0035] (Step S12) In step S12, the first generation unit 12 generates one or more triples by referring to the query. A more detailed explanation of the first generation unit 12 has been given above, so it will be omitted here.

[0036] (Step S13) Next, in step S13, the second generation unit 13 generates a sentence from each of the one or more triples. A more detailed explanation of the second generation unit 13 has been given above, so it will be omitted here.

[0037] (Step S14) Next, in step S14, the first calculation unit 14 obtains a text embedding for each of the generated texts using the trained natural language model, and calculates a score for each of the text embeddings. A more detailed explanation of the first calculation unit 14 has been given above, so it will be omitted here.

[0038] (Step S15) Next, in step S15, the second calculation unit 15 calculates the score for each of the one or more triples' knowledge graph embeddings. A more detailed explanation of the second calculation unit 15 has been given above, so it will be omitted here.

[0039] (Step S16) Next, in step S16, the aggregation unit 16 aggregates the score calculated by the first calculation unit 14 and the score calculated by the second calculation unit 15. A more detailed explanation of the aggregation unit 16 has been given above, so it will be omitted here.

[0040] (Effects of information processing method S1) As described above, in the information processing method S1, • Retrieve the query, • Generate one or more triples by referring to the above query, • Generate a sentence from each of the one or more triples mentioned above. For each generated sentence, obtain a sentence embedding using a trained natural language model, and calculate the score for each of those sentence embeddings. • Calculate the score for each of the one or more triples of the knowledge graph embedding. - The score calculated by the first calculation means and the score calculated by the second calculation means are aggregated. This configuration is employed. According to the above configuration, the same effect as that of the information processing device 1 is achieved.

[0041] (Configuration of Information Processing Device 2) Next, the configuration of the information processing device 2 according to this exemplary embodiment will be described with reference to Figure 4. Figure 4 is a block diagram showing the configuration of the information processing device 2. As shown in Figure 4, the information processing device 2 comprises a first generation unit 21, a second generation unit 22, and a first learning unit 23.

[0042] (First generation unit 21) The first generation unit 21 generates a group of triples including positive example triples and negative example triples by referring to the knowledge graph. Here, it obtains one or more positive example triples by referring to the knowledge graph. Based on these obtained positive example triples, the first generation unit 21 obtains a predetermined number of negative example triples and generates a predetermined number of triple groups including one or more positive example triples and one or more negative example triples. The first generation unit 21 may label positive example triples as positive examples and negative example triples as negative examples. These labels are referenced in the training of the score calculation model SCM described later.

[0043] (Second generation unit 22) The second generation unit 22 generates sentences from each of the triples included in the triple group. For example, for the triples (protein 2, interacts, protein 1) and (protein 2, related, disease 1), the sentences "protein 2 interacts with protein 1" and "protein 2 is related to disease 1" are generated, respectively.

[0044] (Learning Section 1, Part 23) The first learning unit 23 obtains text embeddings for each generated text using a trained natural language model, and trains a scoring model that calculates a score for each of the text embeddings by referring to each of the text embeddings. As an example, the first learning unit 23, • Obtain the vector (text embedding) corresponding to each generated sentence. The above score calculation model is trained by referring to the above vector (text embedding) and the labels (positive or negative examples) attached to the text from which the above vector originated.

[0045] As an example, the first learning unit 23 trains the score calculation model so that the score for positive examples is higher than the score for negative examples. Furthermore, specific examples of trained natural language models are not limited to this exemplary embodiment, but examples include Sentence-BERT and LLM. The natural language model may, for example, acquire domain knowledge by training primarily on literature in the target field. By using such a natural language model for sentence embedding, the generated sentence data is vectorized, and their semantic similarity is quantified. Based on these quantified values, the score calculation model can be trained by referring to each of the sentence embeddings.

[0046] (Effects of Information Processing Device 2) As described above, in the information processing device 2, • Referencing the knowledge graph, generate a group of triples that includes positive and negative triples. • Generate a sentence from each of the triples included in the aforementioned group of triples, For each generated sentence, a sentence embedding obtained using a trained natural language model is acquired, and a scoring model is trained by referencing each of these sentence embeddings to calculate a score for each of those sentence embeddings. This configuration is adopted.

[0047] Thus, in the information processing device 2, • Generate a sentence from each of the one or more triples mentioned above. For each generated sentence, obtain a sentence embedding using a trained natural language model, and calculate a score for each of those sentence embeddings. Because this configuration is adopted, it is possible to suppress increases in costs (time costs and expenses). For example, it is possible to suppress increases in costs compared to a configuration that generates a large number of sentences using a pre-trained natural language model.

[0048] Furthermore, in the information processing device 1, - The score calculated by the first calculation means and the score calculated by the second calculation means are aggregated. Because this configuration is employed, high-precision predictions are possible by utilizing two vector information sources: text embedding and knowledge graph embedding. Therefore, this configuration provides a link (or node) prediction technology that enables lower-cost and higher-precision predictions.

[0049] (Information processing method S2 flow) Next, the flow of the information processing method S2 according to this exemplary embodiment will be explained with reference to Figure 5. Figure 5 is a flowchart showing the flow of the information processing method S2. As shown in Figure 5, the information processing method S2 includes a step (process) S21 of generating a group of triples by referring to a knowledge graph, a step (process) S22 of generating text from the triples, and a step (process) S23 of obtaining text embedding and training a score calculation model by referring to the text embedding.

[0050] (Step S21) In step S21, the first generation unit 21 generates a group of triples including positive example triples and negative example triples by referring to the knowledge graph. A more detailed explanation of the first generation unit 21 has been given above, so it will be omitted here.

[0051] (Step S22) Next, in step S22, the second generation unit 22 generates a sentence from each of the triples included in the triple group. A more detailed explanation of the second generation unit 22 has been given above, so it will be omitted here.

[0052] (Step S23) Next, in step S23, the first learning unit 23 obtains a text embedding for each of the generated texts using the trained natural language model, and trains a scoring model that calculates a score for each of the text embeddings by referring to each of the text embeddings. A more detailed explanation of the first learning unit 23 has been given above, so it will be omitted here.

[0053] (Effects of information processing method S2) As described above, in the information processing method S2, • Referencing the knowledge graph, generate a group of triples that includes positive and negative triples. • Generate a sentence from each of the triples included in the aforementioned group of triples, For each generated sentence, a sentence embedding obtained using a trained natural language model is acquired, and a scoring model is trained by referencing each of these sentence embeddings to calculate a score for each of those sentence embeddings. This configuration is employed. According to the above configuration, the same effect as that of the information processing device 2 is achieved.

[0054] [Second Embodiment] A second exemplary embodiment, which is an example of an embodiment of the present invention, will be described in detail with reference to the drawings. Components having the same function as those described in the above-described exemplary embodiment are denoted by the same reference numerals, and their descriptions are omitted as appropriate. The scope of application of each technology adopted in this exemplary embodiment is not limited to this exemplary embodiment. That is, each technology adopted in this exemplary embodiment can also be adopted in other exemplary embodiments included in this disclosure, to the extent that no particular technical problems arise. Furthermore, each technology shown in the drawings referenced to describe this exemplary embodiment can also be adopted in other exemplary embodiments included in this disclosure, to the extent that no particular technical problems arise.

[0055] (Configuration of Information Processing System 1A) The configuration of the information processing system 1A according to this exemplary embodiment will be described with reference to Figure 6. Figure 6 is a block diagram showing the configuration of the information processing system 1A. As shown in Figure 6, the information processing system 1A comprises an information processing device 100 and a generation device 50 connected to the information processing device 100 via a network N. Here, the specific configuration of the network N is not limited to this exemplary embodiment, but as an example, a wireless LAN (Local Area Network), a wired LAN, a WAN (Wide Area Network), a public telephone network, a mobile data communication network, or a combination of these networks can be used.

[0056] (Generation device 50) As shown in Figure 6, the generation device 50 comprises a control unit 51, a storage unit 52, and a communication unit 53. The communication unit 53 communicates with devices outside the generation device 50. For example, the communication unit 53 communicates with the information processing device 100 provided by the information processing system 1A. The communication unit 53 transmits data supplied from the control unit 51 to the information processing device 100 and supplies data received from the information processing device 100 to the control unit 51. The data received by the communication unit 53 from the information processing device 100 may include a group of triples, including positive and negative examples, generated by the information processing device 100. Furthermore, the data provided by the communication unit 53 to the information processing device 100 may include at least one of the link data and node data generated by the natural language trained model, described later, based on the group of triples.

[0057] The memory unit 52 stores the Knowledge Graph (KG) trained model KGM and the Natural Language Processing (NLP) trained model LLM. For example, the memory unit 52 stores several parameters that define these models. These parameters are, for example, parameters that have been pre-trained by machine learning (parameters that have undergone update processing by machine learning), but this is not limited to this exemplary embodiment. Furthermore, the natural language trained model is a model that has been trained using sentences generated from triples.

[0058] The control unit 51 obtains the output results from each of the two models by using the two models described above. For example, the control unit 51 inputs the query received from the information processing device 100 into the two language models and obtains the embedded data generated by the two language models. It also provides this data and score to the information processing device 100 via the communication unit 53.

[0059] In this exemplary embodiment, the generation device 50 is illustrated as a separate device from the information processing device 100, but this does not limit this exemplary embodiment. For example, the control unit 51 of the generation device 50, or the function of the natural language model execution unit in the control unit 51, may be provided by the control unit of the information processing device 100. Similarly, two or more models stored in the storage unit 52 of the generation device 50 may be stored in the storage unit of the information processing device 100, and the information processing device 100 itself may be configured to execute those two or more models.

[0060] (Configuration of the information processing device 100) Next, the configuration of the information processing device 100 according to this exemplary embodiment will be described with reference to Figure 6. As shown in Figure 6, the information processing device 100 includes a control unit 10, a storage unit 20, a communication unit 30, and an input / output unit 40. The information processing device 100 has the function of a learning device for training the score calculation model SCM and the aggregation model AM, etc., which will be described later, and the function of an inference device for performing inference processing using the trained score calculation model SCM and the aggregation model AM, etc. The learning processing performed by the information processing device 100 is performed in the learning phase, and the inference processing performed by the information processing device 100 is performed in the inference phase.

[0061] (Communications Section 30) The communication unit 30 communicates with devices outside the information processing device 100. For example, the communication unit 30 communicates with the generation device 50. The communication unit 53 transmits data supplied from the control unit 10 to the generation device 50 and supplies data received from the generation device 50 to the control unit 10. The data transmitted by the communication unit 30 to the generation device 50 may include triples acquired by the control unit 10. The data received by the communication unit 30 from the generation device 50 may include embedded data, score data, etc., generated by the generation device 50 based on the triples.

[0062] (I / O section (output information generation section) 40) The input / output unit 40 is configured to include at least one of the following input / output devices: a keyboard, mouse, display, printer, touch panel, etc. Alternatively, the input / output unit 40 may be configured to have input / output devices such as a keyboard, mouse, display, printer, touch panel, etc. connected to it. In this configuration, the input / output unit 40 receives various types of information from the connected input device to the information processing device 100. The input / output unit 40 also outputs various types of information to the connected output device under the control of the control unit 10. An interface such as USB (Universal Serial Bus) can be used as an input / output unit 40. The input / output unit 40 is sometimes referred to as the output information generation unit.

[0063] (Storage unit 20) The storage unit 20 stores various data referenced by the control unit 10, as well as various data generated by the control unit 10. For example, the storage unit 20 stores: Triple-group TPR • Text Embedding Information SEI • Knowledge Graph (KG) Embedding Information KGEI • Score information SI • Aggregated score information ASI • Prediction results PRED • Query QR • Score calculation model SCM • Integrated AM model These are stored there.

[0064] Here, the triple group TPR may include one or more triples generated by the generation unit 12. The text embedding information SEI includes each component of the text embedding result (text embedding) vector. The knowledge graph embedding information KGEI includes each component of the knowledge graph embedding result (knowledge graph embedding vector). The score information includes the score of the text embedding and the score of the knowledge graph embedding calculated by the calculation unit 14. The aggregated score information ASI includes the score aggregated by the aggregation unit 16. The query QR includes the inquiry (question, instruction) obtained from the user by the acquisition unit 11. The prediction result PERD is information generated by referring to the aggregated score. The score calculation model SCM is a model that calculates the score for each text embedding and includes one or more parameters that are the target of learning (update). The aggregation model AM is a model that calculates the aggregated score and may, for example, include one or more parameters that are the target of learning (update). Details of the data, information, models, etc. stored in the storage unit 20 will be described later.

[0065] (Control Unit 10) As shown in Figure 6, the control unit 10 includes a generation unit 12, a calculation unit 14, an aggregation unit 16, a learning unit 23, and an output unit 27. Here, the generation unit 12 also has the same functions as the first generation unit 12 and second generation unit 13 provided in the information processing device 1 described in Exemplary Embodiment 1, and the first generation unit 21 and second generation unit 22 provided in the information processing device 2 described in Exemplary Embodiment 1, so the generation unit 12 may also be referred to as generation unit 12(13,21,22). Similarly, the calculation unit 14 and the learning unit 23 may also be referred to as calculation unit 14(15) and learning unit 23(24,25,26).

[0066] (Acquisition part 11) The acquisition unit 11 acquires a query QR code input by the user during the inference phase. The acquisition of the query QR code may be performed, for example, via the communication unit 30 or via the input / output unit 40. Here, the query QR code may include, for example, multiple information pieces, such as information pieces indicating one or more targets and information pieces indicating the relationship with any of the targets, as in the exemplary embodiment. The acquisition unit 11 extracts these information pieces from the acquired query. For example, if the acquisition unit 11 acquires the query "Which protein is highly associated with disease 1?", it extracts "disease 1" as the information piece indicating the target and "highly associated" as the information piece indicating the relationship with the target. The acquisition unit may also further extract "protein" as the information piece indicating the target.

[0067] The query QR acquired by the acquisition unit 11 may be expressed using knowledge graph terminology as follows: An example of such a query acquired by the acquisition unit 11 is a sentence where one of the components of a triple is missing, or a sentence where one of the components of a triple is relatively high in abstraction. For example, in the above example, "protein" is relatively high in abstraction, and there are various concrete examples of protein; therefore, such concrete examples can be the answer to the query QR.

[0068] An example of a query QR acquired by the acquisition unit 11 is as shown in the example above. In addition to descriptions such as "Which proteins are highly associated with disease 1?" for node prediction, • This may include descriptions such as "What is the relationship between disease 2 and protein 1?" to perform link prediction.

[0069] The format of the above query QR is not particularly limited and may be written in a format that can be applied to any given format, such as listing only the head and tail of the triple, or only the head and relation, or the format may be changed as appropriate. The query should include information on at least one of the following: regimen, drug, gene, and disease. (Generation unit 12) The first generation unit 21 generates triple group TPRs (Total Proportional Representations) that include positive example triples and negative example triples by referring to the knowledge graph during the learning phase. Here, it obtains one or more positive example triples by referring to the knowledge graph. Based on these obtained positive example triples, the generation unit 12 obtains a predetermined number of negative example triples during the learning phase and generates a predetermined number of triple group TPRs that include one or more positive example triples and one or more negative example triples. The generation unit 12 may label positive example triples as positive examples and negative example triples as negative examples. These labels are referenced in the learning of the score calculation model described later.

[0070] Here, each triple in the triple group TPR is defined by a link as a component of the knowledge graph and the nodes at both ends of the link, similar to the example embodiment 1. For example, if the knowledge graph is configured as a directed graph, the triple is composed of (a node as the head, a link as the relation, and a node as the tail). The knowledge graph and triples were described in example embodiment 1 with reference to Figure 3.

[0071] The smallest record unit that makes up a knowledge graph is a triplicate (hereinafter referred to as a triple) representing the constituent elements and their relationships, and a knowledge graph is constructed by enumerating these triples. The constituent elements at both ends (head, tail) correspond to the nodes mentioned above, and the relationship in the middle corresponds to the links mentioned above. The order within the triples in Figure 3 is indicated by the direction of the arrows.

[0072] An example of a triple-group TPR generated by the generation unit 12 during the learning phase in this exemplary embodiment is shown in the upper part of Figure 10. Here, "drug 1", "drug 2", and "protein 2" correspond to the head, "prescribe", "interact", and "relation" correspond to the relation, and "disease 1" and "protein 1" correspond to the tail.

[0073] Furthermore, in the inference phase, the generation unit 12 generates one or more triples by referring to the query QR acquired by the acquisition unit 11. Here, the one or more triples generated by the generation unit 12 in the inference phase are sometimes also referred to as triple group TPRs. However, triple group TPRs generated in the inference phase and triple group TPRs generated in the learning phase may generally be different. Triple group TPRs generated in the inference phase are sometimes denoted as triple group TPRIs, and triple group TPRs generated in the learning phase are sometimes denoted as TPRLs.

[0074] The second generation unit 22 generates sentences from each of the triples included in the triple group TPR. For example, in the inference phase, the second generation unit 22 generates one or more sentences by referring to the triple group TPRI generated by the generation unit 12, and in the learning phase, it generates one or more sentences by referring to the triple group TPRL generated by the generation unit 12. Here, for example, as shown in the bottom two rows of the bottom row of Figure 10, for the triples (protein 2, interacts, protein 1) and (protein 2, related, disease 1), the sentences "protein 2 interacts with protein 1" and "protein 2 is related to disease 1" are generated, respectively.

[0075] (Calculation section 14) In the inference phase, the calculation unit 14 obtains text embeddings for each of the one or more texts generated by the generation unit 12 using a trained natural language model (for example, the NLP trained model LLM described above), and calculates the score of each text embedding by inputting each of the text embeddings into a machine learning-trained score calculation model SCM. Here, "text embedding" refers, as in the exemplary embodiment 1, to, for example, the vector corresponding to the text (vectorizing the text). Here, the score of each text embedding can be calculated by inputting each of the text embeddings into the score calculation model SCM trained by the first learning unit 23, which will be described later.

[0076] The calculation unit 14 calculates the knowledge graph embedding for each of the one or more triples, and calculates the knowledge graph embedding score by performing link prediction for each of the one or more triples by referring to the calculated knowledge graph embedding. Knowledge graph embedding is a method of embedding the components (triples) of the knowledge graph as vectors, and the score is calculated based on the vector values ​​obtained here. In other words, "calculating the knowledge graph embedding score of a triple" means, for example, vectorizing the triple and calculating the score of the vector. The knowledge graph embedding process can be performed, for example, using a trained embedding model (the KG embedding model KGM mentioned above).

[0077] Link prediction is a method of analyzing knowledge graphs that predicts possible links that are not currently present in the data. For example, in Figure 9, it predicts an unknown link that causes disease 2 (is protein 2 involved?). Alternatively, in addition to link prediction, it may also predict nodes connected to a given node by a specified link. Specifically, in Figure 9, it predicts an unknown node that improves disease 2.

[0078] (Learning Section 23) In the learning phase, the learning unit 23 obtains text embeddings for each of the sentences generated by the generation unit 12 using a trained natural language model (for example, the NLP trained model LLM mentioned above), and trains a score calculation model SCM, which calculates a score for each of the text embeddings, by referring to each of the text embeddings. For example, the learning unit 23, • Obtain the vector (text embedding) corresponding to each generated sentence. The above score calculation model is trained by referring to the above vector (text embedding) and the labels (positive or negative examples) attached to the text from which the above vector originated.

[0079] As an example, the learning unit 23 trains the score calculation model so that the score for positive examples is higher than the score for negative examples. Furthermore, specific examples of trained natural language models are not limited to this exemplary embodiment, but examples include Sentence-BERT and LLM. The natural language model may, for example, acquire domain knowledge by learning primarily from literature in the target field, or it may be a model trained by machine learning using training data in the medical or biochemical fields.

[0080] This type of natural language model-based text embedding vectorizes the generated text data and quantifies their semantic similarity. Based on these quantified values, the score calculation model can learn by referencing each of the text embeddings. The obtained text embeddings are stored in the memory unit 52 as text embedding information SEI. Based on this text embedding information SEI, the score calculation model can learn by referencing each of the text embeddings.

[0081] Furthermore, in the learning phase, the learning unit 23 trains an embedding model (the KG embedding model KGM described above) that performs knowledge graph embedding by referring to each of the triples included in the triple group TPR. As described above, knowledge graph embedding is a method of embedding the components (triples) of the knowledge graph as vectors, and here, these vectors are learned. The learned vectors are stored in the storage unit 52 as KG embedding information KGEI.

[0082] As an example, the learning unit 23 may, during the learning phase, train the embedding model KGM by referring to the labels (positive or negative examples) attached to the sentences that were the source of the vectors. Alternatively, as an example, the learning unit 23 may train the embedding model KGM such that the score for positive examples is higher than the score for negative examples.

[0083] Furthermore, in the learning phase, the learning unit 23 trains a link prediction model that performs link prediction by referring to the knowledge graph embedding, by referring to each of the triples included in the triple group TPR. Here, link prediction is a method of analyzing the knowledge graph and predicting links that are not included in the data but could exist. For example, in Figure 9, an unknown link that causes disease 2 is predicted (is protein 2 involved?). Alternatively, in addition to link prediction, nodes connected to a certain node by a specified link may also be predicted. Specifically, in Figure 9, an unknown node that improves disease 2 is predicted.

[0084] The link and node prediction according to this exemplary embodiment is a framework for predicting missing links or missing nodes in a knowledge graph, and can integrate and utilize multiple feature formats, such as latent (learned) features, relational features, and numerical features. These features can have a large number of real values.

[0085] In the exemplary link and node predictions, various combinations of feature formats are possible by integrating embedding-based learning into a probabilistic model in two ways, for example. The first method enables the integration of these feature formats into an end-to-end differentiable learning system by modeling numerical features with radially predetermined functions. The second method combines feature formats using a probabilistic expert product (PoE) approach. For each relationship, there are separate experts that handle different feature formats. The scores of the three submodels (latest variable expert, relational feature expert, and numerical feature expert) are then added together and normalized.

[0086] The link and node predictions in the exemplary embodiments can perform combinations of different feature expression formulas and are knowledge-based learning approaches that combine numerical features, relational features, and latent features. They exhibit superior accuracy on a set of benchmark datasets and can also extract interpretable rule-based features. By presenting these rules to the user, it is possible to explain the link predictions and enable the user to perform visual and temporal tasks. Furthermore, the present invention may be implemented not only in the biomedical field such as diseases, proteins, and drugs, but also as general link or node predictions such as human correlations, human-object relationships, and object-material relationships.

[0087] Furthermore, during the learning phase, the learning unit 23 trains an aggregation model AM that calculates an aggregated score from the score calculated by the score calculation model SCM and the score calculated by the link prediction model. The aggregated score is stored in the storage unit 52 as aggregated score information ASI. The two scores before aggregation may also be stored in the storage unit 52 as score SI.

[0088] In a model that performs a weighted average as an example of an aggregation model, the coefficients of the said weighted average are: • Results of Knowledge Graph Embedding • Results of text embedding • Positive and negative example labels attached to the original triple The configuration may involve determining (updating the coefficients) by referring to at least one of the following. Furthermore, the aggregation model may be a simple average model rather than a weighted average model.

[0089] (Aggregation section 16) The aggregation unit 16 aggregates the text embedding score calculated by the calculation unit 14 and the knowledge graph embedding score. This aggregation allows for highly accurate scoring (i.e., prediction) of the validity of the existence of each triple. The aggregation unit 16 also generates a prediction result PRED by referring to the aggregated score. This prediction result PRED may include the aggregated score. The prediction result PRED from the aggregation unit 16 may be visually presented to the user via a display unit (not shown) or provided to other devices via a communication unit (not shown).

[0090] (Output section 27) The output unit 27 presents the prediction result PRED derived by the aggregation unit 16 to the user via the output information generation unit 40. For example, the output unit 27 may be configured to visually present the prediction result PRED shown in Figure 11 to the user via the output information generation unit 40. For instance, the output information generation unit 40 may generate output information to support the decision-making of healthcare professionals by referring to the scores aggregated by the aggregation unit 16. (Example of result display) "Link prediction" is equivalent to querying a sentence where the "relation" component of a triple is missing and then filling in the blank. Similarly, "node prediction" is equivalent to retrieving a sentence where the "head" or "tail" component of a triple is missing and then filling in the blank. For example, the node prediction results shown in Figure 11 can be obtained.

[0091] In the example shown in Figure 11, a query QR code, "Which proteins are highly associated with disease 1?", is obtained by the user and retrieved by the retrieval unit 11. The answers (nodes) to this query QR code are displayed as protein 17, protein 6, and so on. These answers correspond to the aggregated scores generated by the aggregation unit 16 described above. In other words, in the example shown in Figure 11, the aggregation unit 16 ranks the nodes associated with the aggregated scores in descending order of aggregated scores (so that higher aggregated scores appear higher), and generates a predicted result PRED that includes the results of this ranking. Such a ranking display can support drug discovery researchers in target exploration during the early phases of the drug discovery process. (Information processing method S3 flow) Next, the flow of the information processing method S3 according to this exemplary embodiment will be described with reference to Figure 7. Figure 7 is a flowchart showing the flow of the information processing method S3 executed by the information processing device 100 according to this exemplary embodiment, and corresponds to the flow in the inference phase. Each step of the information processing method S3 is, for example, the same as each step S11 to S16 of the information processing method S1 according to exemplary embodiment 1, however, in the information processing method S3, steps S13 and S14 can be processed in parallel with step S15. (Step S11) In step S11, the acquisition unit 11 acquires the query QR code. A more detailed explanation of the acquisition unit 11 has been given above, so it will be omitted here.

[0092] (Step S12) In step S12, the generation unit 12 generates one or more triples (triple group TPRI) by referring to the query QR. A more detailed explanation of the generation unit 12 in this step has been given above, so it will be omitted here.

[0093] (Step S13) Next, in step S13, the generation unit 12 generates a sentence from each of the one or more triples. A more detailed explanation of the generation unit 12 in this step has been given above, so it will be omitted here.

[0094] (Step S14) Next, in step S14, the calculation unit 14 obtains a text embedding for each of the generated texts using the trained natural language model, and calculates a score for each of the text embeddings. A more detailed explanation of the calculation unit 14 in this step has been given above, so it will be omitted here.

[0095] (Step S15) Next, in step S15, the calculation unit 14 calculates the score for each of the one or more triples' knowledge graph embeddings. A more detailed explanation of the calculation unit 14 in this step has been given above, so it will be omitted here.

[0096] (Step S16) Next, in step S16, the aggregation unit 16 aggregates the text embedding score calculated by the calculation unit 14 and the knowledge graph embedding score. A more detailed explanation of the aggregation unit 16 has been given above, so it will be omitted here. As described above, the information processing device 200 is A means for obtaining queries (acquisition unit 11), A first generation means (first generation unit 12 (generation unit 12)) that generates one or more triples by referring to the aforementioned query, A second generation means (second generation unit 13 (generation unit 12)) that generates text from each of the one or more triples, For each generated sentence, a first calculation means (first calculation unit 14 (calculation unit 14)) obtains sentence embedding using a trained natural language model and calculates a score for each of the sentence embeddings, A second calculation means (second calculation unit 15 (calculation unit 14)) for calculating the score of each of the one or more triples of the knowledge graph embedding, An aggregation means (aggregation unit 16) that aggregates the score calculated by the first calculation means and the score calculated by the second calculation means. It is equipped with the above-described exemplary embodiment.

[0097] The above configuration provides the same effects as the information processing devices 1 and 2 according to Exemplary Embodiment 1. As a result, for example, it can support target discovery and the testing process of WET trials in the early phases of the drug discovery process, and contribute to reducing the total development cost.

[0098] (Information processing method S4 flow) Next, the flow of the information processing method S3 according to this exemplary embodiment will be explained with reference to Figure 8. Figure 8 is a flowchart showing the flow of the information processing method S4, which corresponds to the flow in the learning phase. As shown in Figure 8, the information processing method S4 includes, in addition to the steps (processes) S21, S22, and S23 described above, a step (process) S24 in which the KG embedding model KGM is trained by referring to a triple group TPR (for example, a triple group TPRL), a step (process) S25 in which the link prediction model is trained by referring to a triple group TPR (for example, a triple group TPRL), and a step (process) S26 in which the score aggregation model AM is trained.

[0099] As shown in this figure, steps S24 and S25 are processed in parallel with steps S22 and S23, but are not limited to this, and may be processed in series, for example.

[0100] (Step S21) In step S21, the generation unit 12 generates a triple group TPRL including positive example triples and negative example triples by referring to the knowledge graph. A more detailed explanation of the generation unit 12 in this step has been given above, so it will be omitted here.

[0101] (Step S22) Next, in step S22, the generation unit 12 generates text from each of the triples included in the triple group TPRL. A more detailed explanation of the generation unit 12 in this step has been given above, so it will be omitted here.

[0102] (Step S23) Next, in step S23, the learning unit 23 obtains a text embedding for each of the generated texts using a trained natural language model, and trains a score calculation model SCM, which calculates a score for each of the text embeddings, by referring to each of the text embeddings. A more detailed explanation of the learning unit 23 in this step has been given above, so it will be omitted here.

[0103] (Step S24) In step S24, the learning unit 23 trains the embedding model KGM, which performs knowledge graph embedding, by referring to each of the triples included in the triple group TPRL. A more detailed explanation of the learning unit 23 in this step has been given above, so it will be omitted here.

[0104] (Step S25) In step S25, the learning unit 23 trains a link prediction model that performs link prediction by referring to the knowledge graph embedding, by referring to each of the triples included in the triple group TPRL. A more detailed explanation of the learning unit 23 in this step has been given above, so it will be omitted here.

[0105] (Step S26) Next, in step S26, the learning unit 23 trains an aggregation model AM that calculates an aggregated score from the score calculated by the score calculation model SCM and the score calculated by the link prediction model.

[0106] As described above, in information processing method S4, • Referencing the knowledge graph, generate a group of triples that includes positive and negative triples. • Generate a sentence from each of the triples included in the aforementioned group of triples, For each generated sentence, a sentence embedding obtained using a trained natural language model is acquired, and a scoring model that calculates a score for each of these sentence embeddings is trained by referring to each of these sentence embeddings. • The embedding model that performs knowledge graph embedding is trained by referencing each of the triples included in the triple group. A link prediction model that performs link prediction by referring to the knowledge graph embedding is trained by referring to each of the triples included in the triple group. • Train an aggregation model that calculates an aggregated score from the scores calculated by the score calculation model and the scores calculated by the link prediction model. This configuration is adopted. According to the above configuration, the same effects as those of the information processing devices 1 and 2 according to exemplary embodiment 1 can be achieved. As a result, for example, it is possible to support target discovery and the testing process of WET trials in the early phases of the drug discovery process, and contribute to reducing the total development cost.

[0107] [Third Embodiment] A third exemplary embodiment, which is an example of an embodiment of the present invention, will be described in detail with reference to the drawings. Components having the same function as those described in the above-described exemplary embodiments are denoted by the same reference numerals, and their descriptions are omitted as appropriate. The scope of application of each technology adopted in this exemplary embodiment is not limited to this exemplary embodiment. That is, each technology adopted in this exemplary embodiment can also be adopted in other exemplary embodiments included in this disclosure, to the extent that no particular technical hindrance occurs. Furthermore, each technology shown in the drawings referenced to describe this exemplary embodiment can also be adopted in other exemplary embodiments included in this disclosure, to the extent that no particular technical hindrance occurs.

[0108] (Configuration of the information processing device 200) The configuration of the information processing device 200 according to this exemplary embodiment will be described with reference to Figure 12. As shown in Figure 12, the information processing device 200 does not have the learning unit 23 among the configurations of the information processing device 100 according to exemplary embodiment 2, but the other configurations are the same as those of the information processing device 100 according to exemplary embodiment 2. In this configuration example, the information processing device 200 is connected to a WET testing device 60 that performs WET testing via a network N. Here, the term WET testing refers to testing that actually uses proteins, reagents, etc., and is used in contrast to DRY testing such as data analysis. In other words, the WET testing device 60 is a device that performs testing that actually uses proteins, reagents, etc.

[0109] The output information generation unit (input / output unit) 40 of the information processing device 200 may refer to the prediction result PRED generated by the aggregation unit 16 and generate instruction information indicating the test content that the WET test device 60 should perform. The output information generation unit 40 then transmits the generated instruction information to the WET test device 60 via the network N. The WET test device 60 then refers to the instruction information provided by the information processing device 200 and performs the test indicated by the instruction information. For example, if the prediction information generated by the output information generation unit 40 is the prediction result PRED shown in Figure 11, the output information generation unit 40 may instruct the WET test device 60 to perform a test using protein 17, which is ranked highest. Alternatively, the output information generation unit 40 may assign a priority that has a positive correlation with the aggregated score (0.7, 0.4, ... in Figure 11) to the corresponding nodes (protein 17, protein 6, ...) and instruct the WET test device to perform the tests in the order corresponding to that priority. Such instructions can, for example, support evaluation tests such as WET tests for genome researchers.

[0110] [Examples of implementation using software] Some or all of the functions of the information processing devices 1, 2, 100, and 200 (hereinafter also referred to as "each of the above devices") may be implemented by hardware such as integrated circuits (IC chips) or by software.

[0111] In the latter case, each of the above devices is implemented, for example, by a computer that executes instructions for a program, which is software that realizes each function. An example of such a computer (hereinafter referred to as Computer C) is shown in Figure 13. Figure 13 is a block diagram showing the hardware configuration of Computer C, which functions as each of the above devices.

[0112] Computer C comprises at least one processor C1 and at least one memory C2. Memory C2 stores a program P that causes computer C to operate as each of the above-mentioned devices. In computer C, processor C1 reads program P from memory C2 and executes it, thereby realizing each of the above-mentioned devices.

[0113] For processor C1, for example, a CPU (Central Processing Unit), GPU (Graphic Processing Unit), DSP (Digital Signal Processor), MPU (Micro Processing Unit), FPU (Floating Point Number Processing Unit), PPU (Physics Processing Unit), TPU (Tensor Processing Unit), quantum processor, microcontroller, or a combination thereof can be used. For memory C2, for example, flash memory, HDD (Hard Disk Drive), SSD (Solid State Drive), or a combination thereof can be used.

[0114] Computer C may also be equipped with RAM (Random Access Memory) for loading program P at runtime and for temporarily storing various data. Furthermore, computer C may be equipped with communication interfaces for sending and receiving data with other devices. Additionally, computer C may be equipped with input / output interfaces for connecting input / output devices such as keyboards, mice, displays, and printers.

[0115] Furthermore, program P can be recorded on a non-temporary, tangible recording medium M that is readable by computer C. Such a recording medium M could be, for example, tape, disk, card, semiconductor memory, or programmable logic circuitry. Computer C can acquire program P via such a recording medium M. Program P can also be transmitted via a transmission medium. Such a transmission medium could be, for example, a communication network or broadcast waves. Computer C can also acquire program P via such a transmission medium.

[0116] [Additional Note A] This disclosure includes the technologies described in the following appendices. However, the present invention is not limited to the technologies described in the following appendices, and various modifications are possible within the scope of the claims.

[0117] (Note A1) The means of obtaining the query, A first generation means that generates one or more triples by referring to the aforementioned query, A second generation means for generating text from each of the one or more triples, For each generated sentence, a first calculation means obtains a sentence embedding using a trained natural language model and calculates a score for each of the sentence embeddings. A second calculation means for calculating the score of each of the one or more triples of the knowledge graph embedding, An aggregation means that aggregates the score calculated by the first calculation means and the score calculated by the second calculation means. An information processing device equipped with the following features.

[0118] (Appendix A2) The first calculation means is, The score for each of the aforementioned text embeddings is calculated by inputting it into a machine learning-based scoring model. The information processing device described in Appendix A1.

[0119] (Note A3) The second calculation means is, The knowledge graph embedding for each of the one or more triples is calculated. The score is calculated by performing link predictions for each of the one or more triples, referencing the calculated knowledge graph embedding. The information processing device described in Appendix A2.

[0120] (Note A4) The aforementioned natural language model is a machine learning model that references training data in the medical or biochemical fields. An information processing device as described in any one of the appendices A1 to A3.

[0121] (Note A5) The aforementioned query includes information on at least one of the following: regimens, drugs, genes, and diseases. An information processing device as described in any one of the appendices A1 to A3.

[0122] (Note A6) The system further includes an output information generation means that generates output information to support decision-making by healthcare professionals by referring to the scores aggregated by the aggregation means. An information processing device as described in any one of the appendices A1 to A3.

[0123] (Note A7) A first generation means that generates a group of triples including positive example triples and negative example triples by referring to a knowledge graph, A second generation means for generating a sentence from each of the triples included in the triple group, A first learning means that, for each generated sentence, obtains a sentence embedding using a trained natural language model, and trains a scoring model that calculates a score for each of the sentence embeddings by referring to each of the sentence embeddings. An information processing device equipped with the following features.

[0124] (Note A8) A second learning means for training an embedding model that performs knowledge graph embedding by referencing each of the triples included in the triple group, A third learning means for training a link prediction model that performs link prediction by referring to the knowledge graph embedding by referring to each of the triples included in the triple group, It is equipped with The information processing device described in Appendix A7.

[0125] (Note A9) A fourth learning means for training an aggregation model that calculates an aggregated score from the score calculated by the score calculation model and the score calculated by the link prediction model. An information processing device described in Appendix A8, which is equipped with the following:

[0126] [Additional Note B] This disclosure includes the technologies described in the following appendices. However, the present invention is not limited to the technologies described in the following appendices, and various modifications are possible within the scope of the claims.

[0127] (Note B1) At least one processor handles the retrieval process to obtain the query, The at least one processor performs a first generation process that generates one or more triples by referring to the query, The at least one processor performs a second generation process which generates a document from each of the one or more triples, The at least one processor performs a first calculation process in which it obtains a text embedding for each of the generated texts using a trained natural language model and calculates a score for each of the text embeddings. The at least one processor performs a second calculation process which calculates the score of each of the one or more triples of the knowledge graph embedding, The at least one processor performs an aggregation process that aggregates the score calculated by the first calculation process and the score calculated by the second calculation process. An information processing method that includes this.

[0128] (Note B2) The first calculation process described above is: The score for each of the aforementioned text embeddings is calculated by inputting it into a machine learning-based scoring model. The information processing method described in Appendix B1.

[0129] (Note B3) The second calculation process described above is: The knowledge graph embedding for each of the one or more triples is calculated. The score is calculated by performing link predictions for each of the one or more triples, referencing the calculated knowledge graph embedding. The information processing method described in Appendix B2.

[0130] (Note B4) The aforementioned natural language model is a machine learning model that references training data in the medical or biochemical fields. The information processing method described in any one of the appendices B1 to B3.

[0131] (Note B5) The aforementioned query includes information on at least one of the following: regimens, drugs, genes, and diseases. The information processing method described in any one of the appendices B1 to B3.

[0132] (Note B6) The at least one processor further includes an output information generation process that generates output information to support decision-making by healthcare professionals by referring to the scores aggregated by the aggregation process. The information processing method described in any one of the appendices B1 to B3.

[0133] (Note B7) The at least one processor performs a first generation process that generates a group of triples including positive example triples and negative example triples by referring to a knowledge graph, The at least one processor performs a second generation process which generates a document from each of the triples included in the triple group, The first learning process involves the at least one processor obtaining a text embedding for each of the generated texts using a trained natural language model, and training a scoring model that calculates a score for each of the text embeddings by referring to each of the text embeddings. An information processing method that includes this.

[0134] (Note B8) The at least one processor performs a second learning process in which it trains an embedding model that performs knowledge graph embedding by referencing each of the triples included in the triple group, The at least one processor performs a third learning process in which it trains a link prediction model that performs link prediction referencing the knowledge graph embedding by referencing each of the triples included in the triple group, Includes The information processing method described in Appendix B7.

[0135] (Note B9) A fourth learning process in which at least one processor trains an aggregation model that calculates an aggregated score from the scores calculated by the score calculation model and the scores calculated by the link prediction model. The information processing method described in Appendix B8, which includes the information processing method described therein.

[0136] [Additional Note C] This disclosure includes the technologies described in the following appendices. However, the present invention is not limited to the technologies described in the following appendices, and various modifications are possible within the scope of the claims.

[0137] (Note C1) Computers, The means of obtaining the query, A first generation means that generates one or more triples by referring to the aforementioned query, A second generation means for generating text from each of the one or more triples, For each generated sentence, a first calculation means obtains a sentence embedding using a trained natural language model and calculates a score for each of the sentence embeddings. A second calculation means for calculating the score of each of the one or more triples of the knowledge graph embedding, An aggregation means that aggregates the score calculated by the first calculation means and the score calculated by the second calculation means. An information processing program that functions as such.

[0138] (Note C2) The first calculation means is, The score for each of the aforementioned text embeddings is calculated by inputting it into a machine learning-based scoring model. The information processing program described in Appendix C1.

[0139] (Note C3) The second calculation means is, The knowledge graph embedding for each of the one or more triples is calculated. The score is calculated by performing link predictions for each of the one or more triples, referencing the calculated knowledge graph embedding. The information processing program described in Appendix C2.

[0140] (Note C4) The aforementioned natural language model is a machine learning model that references training data in the medical or biochemical fields. An information processing program described in any one of the appendices C1 to C3.

[0141] (Note C5) The aforementioned query includes information on at least one of the following: regimens, drugs, genes, and diseases. An information processing program described in any one of the appendices C1 to C3.

[0142] (Appendix C6) Computers, The aggregation means further functions as an output information generation means, which generates output information to support decision-making by healthcare professionals by referring to the aggregated scores. An information processing program described in any one of the appendices C1 to C3.

[0143] (Note C7) Computers, A first generation means that generates a group of triples including positive example triples and negative example triples by referring to a knowledge graph, A second generation means for generating a sentence from each of the triples included in the triple group, A first learning means that, for each generated sentence, obtains a sentence embedding using a trained natural language model, and trains a scoring model that calculates a score for each of the sentence embeddings by referring to each of the sentence embeddings. An information processing program that functions as such.

[0144] (Note C8) The aforementioned computer, A second learning means for training an embedding model that performs knowledge graph embedding by referencing each of the triples included in the triple group, A third learning means for training a link prediction model that performs link prediction by referring to the knowledge graph embedding by referring to each of the triples included in the triple group, To make it function as The information processing program described in Appendix C7.

[0145] (Note C9) The aforementioned computer, A fourth learning means for training an aggregation model that calculates an aggregated score from the score calculated by the score calculation model and the score calculated by the link prediction model. To make it function as The information processing program described in Appendix C8.

[0146] [Additional Note D] This disclosure includes the technologies described in the following appendices. However, the present invention is not limited to the technologies described in the following appendices, and various modifications are possible within the scope of the claims.

[0147] (Note D1) The process of retrieving queries, A first generation process that generates one or more triples by referring to the aforementioned query, A second generation process that generates text from each of the one or more triples, For each generated sentence, a first calculation process is performed to obtain a sentence embedding using a trained natural language model and to calculate a score for each of the sentence embeddings. A second calculation process for calculating the score of each of the one or more triples of the knowledge graph embedding, An aggregation process that aggregates the score calculated by the first calculation process and the score calculated by the second calculation process. An information processing device that performs the following actions. The information processing device may also include memory. Furthermore, the memory may store a program that causes at least one processor to execute each of the aforementioned processes.

[0148] (Note D2) The first calculation process described above is: The score for each of the aforementioned text embeddings is calculated by inputting it into a machine learning-based scoring model. The information processing device described in Appendix D1.

[0149] (Note D3) The second calculation process described above is: The knowledge graph embedding for each of the one or more triples is calculated. The score is calculated by performing link predictions for each of the one or more triples, referencing the calculated knowledge graph embedding. The information processing device described in Appendix D2.

[0150] (Note D4) The aforementioned natural language model is a machine learning model that references training data in the medical or biochemical fields. An information processing device as described in any one of the appendices D1 to D3.

[0151] (Note D5) The aforementioned query includes information on at least one of the following: regimens, drugs, genes, and diseases. An information processing device as described in any one of the appendices D1 to D3.

[0152] (Note D6) The aforementioned at least one processor, The aggregation process then performs an output information generation process that uses the aggregated scores to generate output information that supports decision-making by healthcare professionals. An information processing device as described in any one of the appendices D1 to D3.

[0153] (Note D7) The aforementioned at least one processor is A first generation process that generates a group of triples including positive example triples and negative example triples by referring to a knowledge graph, A second generation process that generates a sentence from each of the triples included in the aforementioned triple group, For each generated sentence, a first learning process involves obtaining a sentence embedding using a trained natural language model, and training a scoring model that calculates a score for each of the sentence embeddings by referring to each of the sentence embeddings. An information processing device that performs the following actions. (Note D8) The aforementioned at least one processor is A second learning process involves training an embedding model that performs knowledge graph embedding by referencing each of the triples included in the triple group, A third learning process involves training a link prediction model that performs link prediction by referring to the aforementioned knowledge graph embedding, by referring to each of the triples included in the triple group, Execute The information processing device described in Appendix D7.

[0154] (Note D9) The aforementioned at least one processor is A fourth learning process involves training an aggregation model that calculates an aggregated score from the scores calculated by the score calculation model and the scores calculated by the link prediction model. The information processing device described in Appendix D8 that performs the following actions.

[0155] [Additional Note E] This disclosure includes the technologies described in the following appendices. However, the present invention is not limited to the technologies described in the following appendices, and various modifications are possible within the scope of the claims.

[0156] (Note E1) The process of retrieving queries, A first generation process that generates one or more triples by referring to the aforementioned query, A second generation process that generates text from each of the one or more triples, For each generated sentence, a first calculation process is performed to obtain a sentence embedding using a trained natural language model and to calculate a score for each of the sentence embeddings. A second calculation process for calculating the score of each of the one or more triples of the knowledge graph embedding, An aggregation process that aggregates the score calculated by the first calculation process and the score calculated by the second calculation process. A non-temporary recording medium that stores an information processing program that executes that program. [Explanation of symbols]

[0157] 1,2,100,200 ···Information Processing Device 11 ... Acquisition unit (acquisition means) 12,21 ···First generation unit (first generation means) (generation unit) 13,22 ···Second generation unit (second generation means) (generation unit) 14 ···First calculation unit (first calculation means) (calculation unit) 15 ···Second calculation unit (second calculation means) (calculation unit) 16. Aggregation unit (aggregation means) 23 ···First Learning Section (Learning Method) (Learning Section)

Claims

1. The means of obtaining the query, A first generation means that generates one or more triples by referring to the query, A second generation means for generating text from each of the one or more triples, For each generated sentence, a first calculation means obtains a sentence embedding using a trained natural language model and calculates a score for each of the sentence embeddings. A second calculation means for calculating the score of each of the one or more triples of the knowledge graph embedding, An aggregation means for aggregating the score calculated by the first calculation means and the score calculated by the second calculation means. An information processing device equipped with the following features.

2. The first calculation means is, The score for each of the aforementioned text embeddings is calculated by inputting it into a machine learning-based scoring model. The information processing apparatus according to claim 1.

3. The second calculation means is, The knowledge graph embedding for each of the one or more triples is calculated. The score is calculated by performing link predictions for each of the one or more triples, referencing the calculated knowledge graph embedding. The information processing apparatus according to claim 2.

4. The aforementioned query includes information on at least one of the following: regimens, drugs, genes, and diseases. An information processing apparatus according to any one of claims 1 to 3.

5. The system further includes an output information generation means that generates output information to support decision-making by healthcare professionals by referring to the scores aggregated by the aggregation means. An information processing apparatus according to any one of claims 1 to 3.

6. A first generation means that generates a group of triples including positive example triples and negative example triples by referring to a knowledge graph, A second generation means for generating a sentence from each of the triples included in the triple group, A first learning means that, for each generated sentence, obtains a sentence embedding using a trained natural language model, and trains a scoring model that calculates a score for each of the sentence embeddings by referring to each of the sentence embeddings. An information processing device equipped with the following features.

7. Retrieving the query, To generate one or more triples by referring to the aforementioned query, To generate a text from each of the one or more triples, For each generated sentence, obtain a sentence embedding using a trained natural language model, and calculate a score for each of those sentence embeddings. Calculate the score for each of the one or more triples of the knowledge graph embedding, The scores of the aforementioned text embedding and the aforementioned knowledge graph embedding are aggregated. An information processing method that includes this.

8. By referring to the knowledge graph, a group of triples including positive and negative triples is generated, To generate a sentence from each of the triples included in the aforementioned group of triples, For each generated sentence, a sentence embedding obtained using a trained natural language model is acquired, and a scoring model that calculates a score for each of these sentence embeddings is trained by referring to each of these sentence embeddings. An information processing method that includes this.

9. A program for causing a computer to function as an information processing device according to claim 1, wherein the program causes the computer to function as the first generation means, the second generation means, the first calculation means, the second calculation means, and the aggregation means.

10. A program for causing a computer to function as an information processing device according to claim 6, wherein the program causes the computer to function as the first generation means, the second generation means, and the first learning means.

Images