Research support system

Abstract

To provide a research support system based on generative AI models, which can be fully utilized in chemical-related fields.SOLUTION: A user selects papers that are thought to be relevant or similar to his / her needs from among numerous published papers and uses the selected papers as learning data. The user looks at a screen of a display device to correct the learning data, relabels parameter relationships, and generates a fine-tuned model. The fine-tuned model is applied to build a research support system.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] One aspect of the present invention relates to a research support system, an information processing system, and an information processing method.

[0002] Note that one embodiment of the present invention is not limited to the above technical field. The technical field of one embodiment of the invention disclosed in this specification relates to an object, a method, or a manufacturing method. Alternatively, one embodiment of the present invention relates to a process, a machine, a manufacture, or a composition of matter. Therefore, more specifically, examples of the technical field of one embodiment of the present invention disclosed in this specification include a semiconductor device, a display device, a light-emitting device, a power storage device, a memory device, a driving method thereof, or a manufacturing method thereof. [Background technology]

[0003] In recent years, there has been active development of language models using neural networks, with large-scale language models (LLMs) attracting particular attention. A large-scale language model is a natural language processing model trained using a large amount of data. A large-scale language model can realize, for example, a dialogue model that responds to user instructions. Non-Patent Document 1 discloses GPT-4 (Generative Pre-trained Transformer 4) (registered trademark) as a large-scale language model, and ChatGPT as a dialogue model.

[0004] The use of large-scale language models has significantly increased the capabilities of natural language processing models. However, as language models become larger, it is difficult to incorporate and operate language models in-house due to the equipment and cost involved. Therefore, one way to use language models is to use external services that provide language models. [Prior art documents] [Patent documents]

[0005] [Patent Document 1] Patent Publication No. 2020-102631 [Non-patent literature]

[0006] [Non-Patent Document 1] Summary of ChatGPT / GPT-4 Research and Perspective Towards the Future of Large Language Models, Yiheng Liu et al. (Submitted on 4 Apr 2023, [online], Internet<URL:https: / / arxiv.org / abs / 2304.01852> Summary of the Invention [Problem to be solved by the invention]

[0007] When attempting to utilize large-scale language models to aid in chemical-related research and development, chemical researchers have difficulty creating generative AI models due to their lack of expertise in generative artificial intelligence (AI), and similarly, engineers specializing in generative AI have difficulty creating generative AI models due to their lack of expertise in the field of chemistry.

[0008] Furthermore, while generative AI can generate answers to a variety of questions, when it generates answers related to information it has not learned, it may produce hallucinations, generating answers that are different from the actual answers. In particular, terms related to complex and specialized chemical fields (organic or inorganic chemistry) may not be in the database due to the lack of data, making them more susceptible to hallucinations.

[0009] Therefore, one of our goals is to provide a research support system that uses a generative AI model that can be fully utilized in chemistry-related fields. [Means for solving the problem]

[0010] Since it is difficult to obtain answers (specifically, molecular information) to requests or questions in highly specialized fields (specifically, chemistry) using only large-scale language models that are generally available, the user first extracts papers that they believe are related or similar to the user's request from the many published papers and uses these as training data. The papers used for training contain at least molecular information.

[0011] In addition, user requests are also used as learning data.

[0012] It is preferable to digitize the text data or molecular information in order to introduce it into the model. A known model (JT-VAE (Junction Tree Variational Autoencoder)) is used as the model for digitizing the molecular information. JT-VAE is one of the generative models for molecular structures that applies unsupervised learning VAE to molecular structures. The JT-VAE model is trained in advance using a publicly available database containing a large number of molecular structures. A known model (embedding) is used as the model for digitizing the text data.

[0013] These models are used to generate a fine-tuned model. Then, mapping information (labeling data) is generated regarding the similarity that is output when a request sentence is input to the fine-tuned model. This mapping information is displayed as an image on a display device so that the user can visualize it.

[0014] For example, a mark representing the request statement is placed in the center of the screen, and a mapping display is created showing the relationship between the request statement and data with similarities.

[0015] The user can view the display screen of the display device, correct or add to the data, and relabel the associations of the parameters (similar or dissimilar) to construct an optimal model (fine-tuned model).

[0016] The research support system disclosed in this specification accepts user request statement data and multiple paper data containing sentences from a user device, divides the input data into multiple pieces of sentence data, processes the sentence data using an embedded model to output numerical values as explanatory variables, places the output data in a vector space, outputs a data distribution diagram centered on the user request statement data to the display screen of the user device, and estimates data corresponding to the user request statement data using a fine-tuned model that has been trained by modifying the parameters of the data distribution diagram displayed on the user device based on user operations.

[0017] In the above configuration, the user's request sentence data is part of the training data for the fine-tuned model.

[0018] In the above configuration, the paper data is part of the training data for the fine-tuned model.

[0019] In the above configuration, the modification based on the user's operation modifies the parameter value of the similarity of data similar to the user's request sentence data, or adds new sentence data and parameter values that the user thinks are highly related.

[0020] In this specification, the model is created by, for example, AI, particularly machine learning or deep learning, and is created using, for example, a neural network, a support vector machine, a random forest, or the like. [Effects of the Invention]

[0021] A research support system can be realized that uses a language model to provide molecular structures corresponding to user requests. [Brief explanation of the drawings]

[0022] [Figure 1] FIG. 1 is a block diagram showing a schematic configuration of a research support system according to one embodiment of the present invention. [Figure 2]FIG. 2 is an example of a flowchart showing one embodiment of the present invention. [Figure 3] FIG. 3 is a schematic diagram showing an example of the configuration of an information processing system according to an embodiment of the present invention. [Figure 4] FIG. 4 is an example of a flowchart showing one embodiment of the present invention. [Figure 5] FIG. 5 is an example of a block diagram illustrating one embodiment of the present invention. [Figure 6] FIG. 6A is an application example of a flowchart illustrating one embodiment of the present invention, and FIG. 6B is a block diagram illustrating one embodiment of the present invention. [Figure 7] FIG. 7 is a schematic diagram showing an example of the configuration of an information processing system according to one embodiment of the present invention. [Figure 8] FIG. 8 is a diagram showing a part of the mapping information output from the model. [Figure 9] FIG. 9 shows an example of a display image before correction by a user operation. [Figure 10] FIG. 10 shows an example of a displayed image after correction by a user operation. DETAILED DESCRIPTION OF THE INVENTION

[0023] Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings. However, the present invention is not limited to the following description, and it will be readily understood by those skilled in the art that various modifications can be made to the embodiments and details. Furthermore, the present invention should not be interpreted as being limited to the description of the embodiments shown below.

[0024] (Embodiment 1) Figure 1 shows a block diagram for creating a pre-trained model, which is one of the components of a research support system.

[0025] An example of creating a fine-tuned model 101 based on model 100 is shown below.

[0026] Fine tuning is performed using a display device 150. Specifically, the display device 150 is a personal computer or an information processing terminal with a touch panel that allows touch input. The user modifies data (setting similarity parameter values) or adds data while viewing the image displayed on the display device 150. The image displayed on the display device 150 is a mapping image that visualizes the output data of the model 100, and is preferably an easy-to-understand image for the user. The mapping image is created based on the annotation information, with additional information, such as annotations, added. Specifically, the user's desired item, a request statement 200, or a desired parameter, is placed as a single mark in the center of the screen, and other parameters (statements) are placed according to the distance to the mark, indicating the degree of similarity. Figure 1 shows an example in which a mark indicated by a white frame is placed in the center of the image displayed on the display device 150. Furthermore, parameters with the highest similarity are linked together. The closer the distance from the center of the screen, the higher the similarity, and the target item and related phrases are displayed with lines connecting them.

[0027] If the user creates a modified display image by making modifications, a fine-tuned model 101 that reflects the modified image can be created. The user uses the display device to relabel the relevance of parameters (similar or dissimilar). In this way, the fine-tuned model 101 is created based on the user's operations.

[0028] The training data for the model 100 is created in advance by a user researcher using published papers 130 or request statements 200. The user researcher inputs the target item. The target item may be, for example, a request item or a research topic, such as the item name "reliability of materials." The user researcher inputs which parameters to use as variables, and data related to the sentences or terms in the paper 130 is collected using a large-scale language model, and annotation processing is performed on that data.

[0029] The annotation process involves adding annotations, which are additional information, and can be associated with the positional relationship of the mapping image.

[0030] The model 100 uses an embedded model. In this embodiment, a mapping (network) image is created using a model called multilingual-e5. Character strings are extracted for each period using the model 100, and the similarity of each character string is output using the embedded model.

[0031] Furthermore, the data input to the model 100 is not limited to papers 130, but may also include experimental data and measurement data actually obtained by the user, and may also utilize data from publicly available databases, specifically molecular information, patents, chemical technology journals, etc. By using specialized information as learning data, the output result of the model 100, i.e., the optimal solution, can be obtained.

[0032] Furthermore, by using the fine-tuned model 101 created based on the user's operation, an optimal solution can be obtained with high accuracy.

[0033] FIG. 2 shows an example of a flowchart illustrating a routine for creating a fine-tuned model 101 according to an embodiment of the present invention.

[0034] Referring to FIG. 2, in step S101, article data and request statement data are input to the user device 10. In the following step S102, data is transmitted from the user device 10 to the first server 11, and the article data and request statement data are converted into vectors in the first server 11. Here, the vector refers to data obtained by encoding explanatory text. In the following step S103, the user device 10 visualizes the relevance between sentences (constructs a sentence network) using the similarity of each vector, and modifies the network based on user operations. Contrastive Loss or Multiple Negatives Ranking Loss can be used as a specific method for reflecting user modifications in the fine-tuned model 101.

[0035] In the following step S104, the first server 11 creates a fine-tuned model 101 that reflects the correction.

[0036] FIG. 3 shows an example of the configuration of each component of an information processing system according to one embodiment of the present invention, a database, and a network connecting these components.

[0037] As shown in FIG. 3, an information processing system (also referred to as a research support system) according to one embodiment of the present invention includes at least a user device 10 and a first server 11.

[0038] The user device 10 may be, for example, a desktop computer.

[0039] The user device 10 can accept data input by a user and can provide data output by the first server 11, the second server 12, or the third server 13 to the user.

[0040] For example, dedicated application software or a web browser may be operated. A user can access the information processing system via either of these, thereby enjoying services using the information processing system according to one aspect of the present invention.

[0041] The first server 11 may be, for example, a workstation, a server computer, or a supercomputer. The first server 11 preferably has a function as a parallel computer. By using it as a parallel computer, it is possible to perform large-scale calculations required for AI learning and inference, for example. The first server 11 can perform processing using a natural language processing model using AI.

[0042] For example, it is possible to perform processing (Natural Language Processing) using natural language models such as BERT (Bidirectional Encoder Representations from Transformers), T5 (Text-to-Text Transfer Transformer), GPT-3, GPT-3.5, GPT-4 (registered trademark), LaMDA (Language Model for Dialogue Applications), PaLM (Pathways Language Model), and Llama2.

[0043] The information processing system also exchanges predetermined data between the user device 10 and the first server 11 via a network 50. The information processing system can generate a fine-tuned model 101 by linking the user device 10 and the first server 11.

[0044] (Embodiment 2) 4 shows an example of a flowchart for forming a regression model according to one aspect of the present invention in the fourth server 14. As shown in FIG. 4, an example is shown in which multiple servers (a first server 11, a second server 12, a third server 13, and a fourth server 14) are used.

[0045] Referring to FIG. 4, in step S201, the user selects a paper or document 140 and inputs it into the user device 10. The information input is electronic data including at least a molecular structure and text. In the following step S202, the user device 10 creates an extraction prompt based on the input document data and request data. In the following step S203, the first server 11 extracts and divides the molecular expression and explanatory text. In the following step S204, the user device 10 creates a list.

[0046] In the following step S205, the user device 10 inputs the explanatory sentence to the second server 12. In the following step S206, the second server 12 converts the explanatory sentence into a vector. In the following step S207, the user device 10 inputs the molecular expression to the third server 13. In the following step S208, the third server 13 converts the molecular expression into a vector. In the following step S209, the user device 10 creates a dataset and inputs it to the fourth server 14 as training data. In the following step S210, the regression model 110 is trained so that the molecular vector can be predicted from the sentence vector.

[0047] 5 shows an example of a block diagram of the regression model 110. The regression model 110 is a model before any user operation is performed.

[0048] The training data for the regression model 110 is extracted separately as molecular structure sequences (corresponding to molecular expressions) and related character strings (corresponding to explanatory text) using a large-scale language model (LLM model), and stored in a database 111. Language information other than the document 140 is stored in advance in the database 111. The document 140 and the request statement 200 selected by the user are added as data information to the database 111 using the large-scale language model (LLM model).

[0049] For molecular structures, a model that quantifies molecular structures (typically JT-VAE) is used to output objective variables 112, and for character strings, an embedded model is used to output explanatory variables 113. The more molecular structures contained in the literature 140 prepared by the user, the more corresponding sentences there are for each, and therefore the more training data there is, making it possible to create a regression model 110 with substantial training data.

[0050] 6(A) shows an example of a flowchart illustrating a routine of an application example using the fine-tuned model 101 according to one embodiment of the present invention. Note that since the fine-tuned model 101 is created in the second server 12 using the first server 11, FIG. 6(A) shows an example of creating a new regression model in a fourth server 14, which is another server.

[0051] First, a new regression model is created in order to make the vector space of the model correspond to that of the fine-tuned model 101. The new regression model is created in the fourth server 14.

[0052] Referring to FIG. 6(A), in step S211, the user inputs a wish sentence into the user device 10. In the following step S212, the wish sentence is input into the second server. In the following step S213, the second server 12 converts the wish sentence into a vector. In the following step S214, the fourth server 14 predicts a numerator vector from the sentence vector. In the following step S215, the third server 13 converts the numerator vector into a molecular representation. In the following step S216, the molecular representation is displayed on the user device 10. In the following step S217, the user checks the molecular representation on the display screen of the user device 10.

[0053] FIG. 6B is a schematic diagram of information processing using the fine-tuned model 101, from inputting a request statement 200 to making an estimation and outputting a molecular structure X.

[0054] As shown in FIG. 6(B), when a user inputs a wish statement 200 to the fine-tuned model 101, the fine-tuned model 101 outputs a sequence A corresponding to the wish statement 200. By inputting the sequence A to the regression model 102, a sequence B is output. By further inputting the sequence B to a model (typically JT-VAE), it can be converted into a molecular structure X. The regression model 102 to which the sequence A is input is created to make the vector spaces of the fine-tuned model 101 and the model (JT-VAE) correspond to each other. The user can perform the input, output, and display confirmation shown in FIG. 6(B) using the user device 10 that he or she operates while looking at the display device 150.

[0055] In this way, when a user inputs a request statement 200 into the fine-tuned model 101, a molecular structure X that can serve as a hint for research and development can be obtained. Using such a system as a research support system allows for efficient research and development.

[0056] Furthermore, by varying the request statement 200 in various ways, the user can obtain various molecular structure candidates that can serve as hints for research and development.

[0057] FIG. 7 shows an example of the configuration of each component of an information processing system according to one embodiment of the present invention, various databases, and a network connecting these components.

[0058] In Fig. 7, the same parts as in Fig. 3 are described using the same reference numerals. Fig. 7 shows an example in which a fine-tuned model is constructed in the second server 12, and an example in which a regression model 102 is constructed in the fourth server 14.

[0059] The second server 12, the third server 13, and the fourth server 14 can be, for example, a workstation, a server computer, a supercomputer, etc. The fourth server 14 preferably has a function as a parallel computer. By using it as a parallel computer, it can perform large-scale calculations necessary for AI learning and inference, for example.

[0060] This embodiment mode can be freely combined with other embodiment modes. [Example]

[0061] An embodiment of creating a fine-tuned model will be described below with reference to the drawings.

[0062] In this embodiment, an example of creating a fine-tuned model will be described using Patent Document 1 (JP 2020-102631 A), which has already been published.

[0063] First, data related to the text data of patent documents is extracted. The data is divided so that each sentence is an element. The similarity between each sentence is then calculated and used as an annotation, which is additional information. A mapping (network) image is created using a model called multilingual-e5. Linking is performed when the similarity between a given sentence and another sentence is highest. Similar linking is performed in a brute-force manner according to the total number of sentences. This linking process results in the formation of multiple groups of linked sentences.

[0064] Figure 8 shows multiple groups. Figure 8 shows an example where one sentence is displayed in one block, showing a portion of the full-text data. Note that the user can scroll the screen to check the linking of each sentence to the block of the full-text data.

[0065] Next, the desired sentence 200 is "materials with high color purity." The similarity is calculated for the sentence "materials with high color purity." The calculated results are linked to the top four sentences, and they are arranged in order of similarity so that the sentence "materials with high color purity" is at the center. Table 1 below shows a portion of the results. For simplicity, Table 1 only shows the top seven sentences out of the total sentences.

[0066] [Table 1]

[0067] Next, when determining whether the top four sentences are technically close to "materials with high color purity," the second-place sentence (T2) and the fourth-place sentence (T4) of the top four sentences (T1, T2, T3, T4) are not technically related, so the link is removed.In addition, although it is a technically related sentence, the sentence with the 107th-place similarity (T107) was not in the top four, so it was not linked, but it was linked: "In addition, it was confirmed that the half-width of all of these emission spectra narrowed."

[0068] Figure 9 shows an example of an image before correction, and Figure 10 shows an example of an image after correction. The images shown in Figures 9 and 10 correspond to data distribution maps centered on the request statement (200) data. A fine-tuned model can be created based on the data corresponding to the corrected image.

[0069] By performing the information processing described above, a fine-tuned model was created. [Explanation of symbols]

[0070] 10: User device 11: First server 12: Second server 13: Third server 14: 4th server 100:Model 101: Fine-tuned model 102: Regression Models 110: Regression Models 111: Database 112: Objective variable 113: Explanatory variables 130: Essay 140: Literature 150:Display device 200: Request

Claims

1. A user device receives user request data and a plurality of paper data including sentences, and divides the input data into a plurality of sentence data; The text data is processed using an embedding model to output numerical values as explanatory variables, and the output data is arranged in a vector space; outputting a data distribution chart centered on the user's request statement data on a display screen of the user device; A research support system that estimates data corresponding to the user's request statement data using a fine-tuned model that has been trained by modifying the parameters of the data distribution map displayed on the user device based on user operations.

2. 2. The research support system according to claim 1, wherein the user's request data is part of the training data for the fine-tuned model.

3. 2. The research support system according to claim 1, wherein the paper data is part of the training data for the fine-tuned model.

4. 2. A research support system according to claim 1, wherein the modification modifies the parameter value of the similarity of data similar to the user's request sentence data, or adds new sentence data and parameter values that the user considers highly relevant.

Images