Information processing device, information processing method, and recording medium

Abstract

In this information processing device, a complementation means complements ontology data by using a graph machine learning model that has learned the relationship between information items included in the ontology data. A natural language processing model generation means generates a natural language processing model on the basis of the complemented ontology data. The information processing device can assist decision making of a user.

Description

Information processing device, information processing method, and recording medium

[0001] The present disclosure relates to training natural language processing models.

[0002] Conventionally, graph data such as knowledge graphs that represent relationships between multiple entities have been known. Patent Document 1 proposes a method for establishing new relationships between new entities and existing entities by performing logic-based reasoning and zero-shot learning based on the above-mentioned knowledge graph.

[0003] Japanese Patent Application Laid-Open No. 2019-125364

[0004] In training natural language processing models, knowledge graphs and various natural language documents can be used as training data. In training natural language processing models, it is preferable to use training data that covers all knowledge information in order to improve the training accuracy.

[0005] One object of the present disclosure is to provide an information processing device that complements training data and improves the training accuracy of a natural language processing model.

[0006] In one aspect of the present disclosure, an information processing device includes: a complementing means for complementing ontology data using a graph machine learning model that has learned the relationships between information included in the ontology data; and a natural language processing model generating means for generating a natural language processing model based on the complemented ontology data.

[0007] In another aspect of the present disclosure, an information processing method includes complementing ontology data using a graph machine learning model that has learned relationships between information included in the ontology data, and generating a natural language processing model based on the complemented ontology data.

[0008] In yet another aspect of the present disclosure, a recording medium records a program that causes a computer to execute a process of complementing ontology data using a graph machine learning model that has learned the relationships between information included in the ontology data, and generating a natural language processing model based on the complemented ontology data.

[0009] According to the present disclosure, it is possible to provide an information processing device that complements training data and improves the training accuracy of a natural language processing model.

[0010] 1 is a diagram conceptually illustrating an information processing device; FIG. 2 is a block diagram illustrating a hardware configuration of the information processing device; FIG. 3 is a block diagram illustrating a functional configuration of the information processing device; FIG. 4 is a block diagram illustrating an example of ontology data; FIG. 5 is a graph structure of ontology data; FIG. 6 is a flowchart illustrating an NLP model generation process; FIG. 7 is a block diagram illustrating a functional configuration of an information processing device according to a second embodiment; and FIG. 8 is a flowchart illustrating processing by the information processing device according to the second embodiment.

[0011] Preferred embodiments of the present disclosure will now be described with reference to the drawings. First Embodiment Overall Configuration FIG. 1 is a conceptual diagram of an information processing device. The information processing device 10 trains a natural language processing (hereinafter also referred to as "NLP") model based on input ontology data and natural language documents. In particular, the information processing device 10 of this embodiment is characterized in that it complements the input ontology data, increases the comprehensiveness of the data, and then uses the data for training the NLP model. This makes it possible to train an NLP model specialized in a specific domain, for example, where the comprehensiveness of the input data affects the quality of the output data.

[0012] The specific domain refers to an industry with many specialized terms and examples, such as the medical field or the legal field. In this embodiment, medical ontology / dictionary information is used as ontology data, and information such as medical literature, patient charts, and clinical trial conditions is used as natural language documents.

[0013] 2 is a block diagram showing the hardware configuration of the information processing device 10. As shown in the figure, the information processing device 10 includes a processor 11, an interface (IF) 12, a read-only memory (ROM) 13, a random access memory (RAM) 14, a storage device 15, and an input unit 16. The components are connected to each other via a bus 18, for example.

[0014] The processor 11 is a computer such as a CPU (Central Processing Unit), and executes a program prepared in advance to control the entire information processing device 10. Specifically, the processor 11 may be a CPU, a GPU (Graphics Processing Unit), a DSP (Digital Signal Processor), an MPU (Micro Processing Unit), an FPU (Floating Point number Processing Unit), a PPU (Physics Processing Unit), a TPU (Tensor Processing Unit), a quantum processor, a microcontroller, or a combination thereof.

[0015] The processor 11 also loads programs stored in the ROM 13, the storage device 15, etc., and executes each process coded in the program. The processor 11 also functions as a part or the whole of the information processing device 10. The processor 11 then executes the NLP model generation process described below.

[0016] The IF 12 inputs and outputs data to and from external devices. Specifically, ontology data and natural language documents are input to the information processing device 10 through the IF 12.

[0017] The ROM 13 stores various programs executed by the processor 11. The RAM 14 is used as a working memory while the processor 11 is executing various processes.

[0018] The storage device 15 is a non-volatile, non-transitory storage device such as a disk-shaped recording medium or a semiconductor memory. The storage device 15 may be configured to be detachable from the information processing device 10. The storage device 15 stores various programs executed by the processor 11. The storage device 15 may also store ontology data and natural language documents input from an external device. The storage device 15 may also store a link prediction model, an NLP model, etc., which will be described later.

[0019] The input unit 16 is, for example, a mouse, a keyboard, etc., and is used by the user to input data.

[0020] 3 is a block diagram showing the functional configuration of the information processing device 10 according to the first embodiment. The information processing device 10 functionally includes a link prediction model generation unit 111, a data complementation unit 112, and an NLP model generation unit 113.

[0021] Ontology data is input to the information processing device 10 via the IF 12. The ontology data is input to a link prediction model generation unit 111 and a data complementation unit 112. In addition, a natural language document is input to the information processing device 10 via the IF 12. The natural language document is input to an NLP model generation unit 113.

[0022] An example of ontology data is shown in Fig. 4. Ontology data is configured with a set of three elements, head, relation, and tail (hereinafter also referred to as a "triple"), as its smallest unit.

[0023] FIG. 5 shows ontology data in a graph structure. In this embodiment, of the three-element sets, the head and tail elements are called "nodes," and the relation element is called a "relation." A "relation" is a directional link from the head node to the tail node. For example, in FIG. 5, a directional link is added from the "Higher Classification Disease Name" node to the "Lower Classification Disease Name 1" node, indicating that the relationship is a "subordinate concept." Note that the higher classification disease name and lower classification disease names 1 to 3 in FIG. 5 represent disease names. The higher classification disease name is, for example, a disease name such as lung cancer, and the lower classification disease names 1 to 3 are disease names that further classify the higher classification disease name, for example, lung adenocarcinoma, epithelial lung cancer, non-small cell lung cancer, etc.

[0024] 5 includes general information related to medical care, such as disease name mapping, guideline / drug DB, and academic terms. For example, the disease name mapping is information indicating disease names and the hierarchical relationships between disease names, the guideline / drug DB is information on patient symptoms and information on drugs prescribed to patients, and the academic terms are academic terms contained in medical papers and other literature. The information processing device 10 may also receive ontology data specific to a certain system, in which information on a specific hospital, information on a specific patient group, clinical trial information, and the like is added to the ontology data shown in FIG. 5.

[0025] 3 , the link prediction model generation unit 111 trains a machine learning model using the input ontology data. This machine learning model is trained to estimate relationships between unlinked nodes in the ontology data, and will hereinafter also be referred to as a "link prediction model." The link prediction model generation unit 111 trains the link prediction model by inputting feature amounts corresponding to each node of the ontology data and feature amounts corresponding to each relation into a machine learning model constructed based on, for example, "KBLRN."

[0026] Note that "KBLRN" is disclosed, for example, in Alberto Garcia-Duran, et. al., "KBLRN: End-to-End Learning of Knowledge Base Representations with Latent, Relational, and Numerical Features." Furthermore, the machine learning model may be constructed based on a model other than "KBLRN," as long as it has a configuration capable of performing link prediction for graph data. Furthermore, when training the link prediction model using ontology data, the link prediction model generation unit 111 desirably performs zero-shot learning, such as that disclosed in Patent Document 1.

[0027] The link prediction model generation unit 111 outputs the link prediction model to the data complementation unit 112 .

[0028] Ontology data is input to the data complementing unit 112. In addition, a link prediction model is input to the data complementing unit 112 from the link prediction model generating unit 111. The data complementing unit 112 uses the link prediction model to perform link prediction for some or all of the ontology data, and obtains prediction scores for some or all of the new relationships between nodes. Then, the data complementing unit 112 complements the ontology data based on new relationships whose prediction scores are equal to or greater than a predetermined threshold.

[0029] Specifically, the link prediction model takes a sentence in which one of the three elements of a triple in the input ontology data is missing as a query. The link prediction model then outputs, as a response to the query, completion candidates for the missing element and their prediction scores. The data completion unit 112 then reflects the completion candidates whose prediction scores are equal to or greater than a predetermined threshold in the ontology data, thereby generating completed ontology data.

[0030] Figure 6 shows an example of completed ontology data. In Figure 6, the link prediction model outputs completion candidates with prediction scores above a predetermined threshold from the following queries 1 to 3. Missing elements are indicated by "?". It is assumed that the link prediction model has been trained with the ontology data shown in Figure 5. Query 1: (Subclassification disease name 1, etiology, ?) Output example of completion candidate: ? = "Mutation 1" Query 2: (Subclassification disease name 1, ?, long-term symptoms S) Output example of completion candidate: ? = "Symptoms" Query 3: (?, treatment, treatment 2) Output example of completion candidate: ? = "Subclassification disease name 3"

[0031] The data complementing unit 112 reflects the complement candidates for queries 1 to 3 in the ontology data. Specifically, in Fig. 6, the data complementing unit 112 adds a directional link from "Subclassification Disease Name 1" to "Mutation 1" to the ontology data in Fig. 5, adds a directional link from "Subclassification Disease Name 1" to "Long-term Symptom S", and adds a directional link from "Subclassification Disease Name 3" to "Treatment Method 2".

[0032] In this way, the data complementing unit 112 uses the link prediction model to make predictions for combinations of nodes and relations that exist in the ontology data, and can complement relationships that do not exist in the original ontology data.

[0033] Returning to FIG. 3 , the data complementing unit 112 outputs the complemented ontology data to the NLP model generating unit 113 .

[0034] A natural language document is input to the NLP model generation unit 113. The NLP model generation unit 113 also receives the complemented ontology data from the data complement unit 112. The NLP model generation unit 113 inputs the natural language document and the complemented ontology data to an existing NLP model, and trains the NLP model.

[0035] Existing NLP models include, for example, BERT (Bidirectional Encoder Representations from Transformers), GPT (Generative Pretrained Transformer), T5 (Text-to-Text Transformer), RoBERTa (Robustly optimized BERT approach), and ELECTRA (Efficient Learning an Encoder that Classifies Token Replacements Accurately).

[0036] In the above configuration, the link prediction model generation unit 111 is an example of a learning means, the data complementation unit 112 is an example of a complementation means, and the NLP model generation unit 113 is an example of a natural language processing model generation means.

[0037] [NLP Model Generation Process] Next, the NLP model generation process will be described. Fig. 7 is a flowchart of the NLP model generation process by the information processing device 10. This process is realized by the processor 11 shown in Fig. 2 executing a program prepared in advance and operating as each element shown in Fig. 3.

[0038] First, ontology data and a natural language document are input to the information processing device 10 via the IF 12 (step S111). The ontology data is input to the link prediction model generation unit 111 and the data complementation unit 112. The natural language document is input to the NLP model generation unit 113.

[0039] Next, the link prediction model generation unit 111 uses the input ontology data to train a link prediction model (step S112). The link prediction model is a machine learning model trained to estimate relationships between unlinked nodes in the ontology data. The link prediction model generation unit 111 outputs the link prediction model to the data complementation unit 112.

[0040] Next, the data complementing unit 112 performs link prediction for some or all of the ontology data using the link prediction model, and obtains prediction scores for some or all of the new relationships between nodes.The data complementing unit 112 then complements the ontology data based on the new relationships whose prediction scores are equal to or greater than a predetermined threshold, and generates complemented ontology data (step S113).The data complementing unit 112 outputs the complemented ontology data to the NLP model generation unit 113.

[0041] Next, the NLP model generation unit 113 inputs the natural language document and the completed ontology data into the existing NLP model, and trains the NLP model (step S114). Then, the process ends.

[0042] 8 is a block diagram showing the functional configuration of an information processing apparatus according to Embodiment 2. The information processing apparatus 200 includes a complementing unit 201 and a natural language processing model generating unit 202.

[0043] 9 is a flowchart of processing by the information processing apparatus of the second embodiment. The complementing unit 201 complements the ontology data using a graph machine learning model that has learned the relationships between pieces of information contained in the ontology data (step S201). The natural language processing model generating unit 202 generates a natural language processing model based on the complemented ontology data (step S202).

[0044] According to the information processing device 200 of the second embodiment, it is possible to provide an information processing device that complements training data and improves the training accuracy of a natural language processing model. Furthermore, the information processing device 200 can support user decision-making.

[0045] A part or all of the above-described embodiments can be described as, but not limited to, the following supplementary notes.

[0046] (Supplementary Note 1) An information processing device comprising: a complementing means for complementing ontology data using a graph machine learning model that has learned the relationships between information included in the ontology data; and a natural language processing model generating means for generating a natural language processing model based on the complemented ontology data.

[0047] (Supplementary Note 2) An information processing device according to Supplementary Note 1, comprising a learning means for training the graph machine learning model, wherein the ontology data includes a plurality of nodes and directed links indicating relationships between the nodes, and the learning means trains the graph machine learning model to predict relationships between unlinked nodes in the ontology data using the plurality of nodes included in the ontology data and known relationships between the nodes.

[0048] (Supplementary Note 3) The information processing device according to Supplementary Note 2, wherein the graph machine learning model uses data in which one of three elements, namely, a link source node, a link destination node, and a relationship between the nodes, is missing as a query, and outputs candidates for the missing element and their predicted scores, and the completion means reflects the candidate elements whose predicted scores are equal to or greater than a predetermined threshold in the ontology data.

[0049] (Supplementary Note 4) An information processing device according to Supplementary Note 1, comprising an acquisition means for acquiring ontology data and a natural language document, wherein the natural language processing model generation means generates a natural language processing model based on the complemented ontology data and the natural language document.

[0050] (Appendix 5) An information processing method that uses a graph machine learning model that has learned the relationships between information contained in ontology data to complement the ontology data, and generates a natural language processing model based on the complemented ontology data.

[0051] (Appendix 6) A recording medium that stores a program that causes a computer to execute the process of complementing ontology data using a graph machine learning model that has learned the relationships between information contained in ontology data, and generating a natural language processing model based on the complemented ontology data.

[0052] Although the present disclosure has been described above with reference to the embodiments and examples, the present disclosure is not limited to the above-described embodiments and examples. Various modifications that can be understood by a person skilled in the art can be made to the configuration and details of the present disclosure within the scope of the present disclosure.

[0053] 10 Information processing device 111 Link prediction model generation unit 112 Data complementation unit 113 NLP model generation unit

Claims

1. An information processing device comprising: a complementing means for complementing ontology data using a graph machine learning model that has learned the relationships between information contained in ontology data; and a natural language processing model generation means for generating a natural language processing model based on the complemented ontology data.

2. An information processing device as described in claim 1, comprising a learning means for training the graph machine learning model, wherein the ontology data includes a plurality of nodes and directed links indicating relationships between the nodes, and the learning means trains the graph machine learning model to predict relationships between unlinked nodes in the ontology data using the plurality of nodes included in the ontology data and known relationships between the nodes.

3. The information processing device of claim 2, wherein the graph machine learning model uses data in which one of the three elements, the link source node, the link destination node, and the relationship between the nodes, is missing as a query, and outputs candidates for the missing element and their predicted scores, and the completion means reflects candidate elements whose predicted scores are equal to or greater than a predetermined threshold in ontology data.

4. An information processing device as described in claim 1, further comprising an acquisition means for acquiring ontology data and a natural language document, wherein the natural language processing model generation means generates a natural language processing model based on the complemented ontology data and the natural language document.

5. An information processing method that complements ontology data using a graph machine learning model that has learned the relationships between information contained in ontology data, and generates a natural language processing model based on the complemented ontology data.

6. A recording medium having recorded thereon a program that causes a computer to execute a process of complementing ontology data using a graph machine learning model that has learned the relationships between information contained in ontology data, and generating a natural language processing model based on the complemented ontology data.

Images