Information processing program, information processing method, and information processing device

JPWO2023233633A5Active Publication Date: 2024-12-04FUJITSU LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
JP2024524117
Authority / Receiving Office
JP · JP
Patent Type
Applications
Current Assignee / Owner
Filing Date
2024-09-25
Publication Date
2024-12-04
Estimated Expiration
2042-06-02

AI Technical Summary

Technical Problem

Conventional methods for correcting input errors in sentences, such as typos and omissions, are inadequate in accurately predicting and correcting erroneous sentence vectors, especially when errors arise from incorrect conversions, and they lack effectiveness in improving the accuracy of filling blanks in sentences composed of multiple sentences.

Method used

A computer-based information processing system that calculates vectors for consecutive sentences and uses a machine learning model, trained on pairs of input errors and corrected sentences, to predict and correct sentence vectors by determining the appropriateness of subsequent sentences through cosine similarity analysis, thereby detecting and correcting input errors.

Benefits of technology

The system effectively estimates sentences to fill blanks and detects input errors in sentences composed of multiple sentences, improving the accuracy of error correction and reducing computational costs by generating a machine learning model that predicts and corrects sentence vectors based on trained data.

✦ Generated by Eureka AI based on patent content.
Patent Text Reader

Abstract

This information processing device calculates vectors of a plurality of sentences having a relationship in front and back sentences in a plurality of continuous sentences. The information processing device generates a machine-learning model which predicts sentence vectors of a sentence that is input after a certain sentence when the sentence vectors of the certain sentence are input to the machine-learning model by inputting, in order, the plurality of sentence vectors to the machine-learning model and training the machine-learning model. The information processing device calculates vectors of a first sentence and vectors of a second sentence that follows the first sentence. The information processing device calculates vectors of a sentence that is predicted to follow the first sentence by inputting the vectors of the first sentence to the machine-learning model, and determines whether the vectors of the second sentence are proper.
Need to check novelty before this filing date? Find Prior Art

Description

Information processing program, information processing method, and information processing device

[0001] The present invention relates to an information processing program and the like.

[0002] In recent years, services have been provided that calculate the vector of a sentence and use the calculated vector to perform various processes such as translation into other languages, database search, etc. However, if there is an input error or the like in the sentence itself specified by the user, the vector of the sentence cannot be calculated accurately, and errors may occur in processes such as translation and search.

[0003] For example, the correct sentence "Its function is its characteristic" and the incorrect sentence "Its function is its characteristic" are sentences with significantly different meanings, and the vectors of each sentence are also significantly different.

[0004] A conventional technique for correcting input errors in a sentence involves training a learning model using a data set of pairs of input errors and their corrected sentences from the correction history, and then inputting the target sentence into the trained learning model to correct the input error in the target sentence.

[0005] Japanese Patent Application Laid-Open No. 2019-101993

[0006] Kazuhiro Miki et al., "A Solution to English Blank Fill-in Problems Using BERT," Department of Information Science, Faculty of Engineering, Okayama University, DEIM2020 G2-4 (day1 p47) Tasuku Tanaka et al., "Constructing a Japanese Input Error Dataset Using Wikipedia Revision History," Graduate School of Informatics, Kyoto University, Association for Natural Language Processing, 26th Annual Conference, Proceedings, March 2020

[0007] However, the above-mentioned conventional technology is a technology for filling in gaps in which some words in a sentence are masked, and although it has high accuracy in filling in words that fill in gaps in sentences consisting of multiple words, there are few descriptions of high accuracy in filling in sentences that consist of multiple sentences, and it does not detect sentences that contain input errors. Furthermore, while the conventional technology can correct input errors such as typos and omissions, there are many cases in which it cannot correctly correct input errors resulting from incorrect conversion.

[0008] In one aspect, the present invention aims to provide an information processing program, an information processing method, and an information processing device that can estimate a sentence to fill in a blank in a text consisting of multiple sentences, and detect a sentence that contains an input error.

[0009] In the first proposal, a computer is caused to execute the following process. The computer calculates the vectors of each of a plurality of consecutive sentences that are related to the preceding and following sentences. The computer sequentially inputs the vectors of the plurality of sentences into a machine learning model to train it, thereby generating a machine learning model that, when the vector of a certain sentence is input into the machine learning model, predicts the sentence vector of the sentence that will be input after the given sentence. The computer calculates the vector of a first sentence and the vector of a second sentence that follows the first sentence. The computer inputs the vector of the first sentence into the machine learning model to calculate the vector of the sentence that is predicted to follow the first sentence, and determines whether the vector of the second sentence is appropriate.

[0010] It is possible to estimate sentences to fill in gaps in a text consisting of multiple sentences, and to detect sentences containing input errors.

[0011] FIG. 1 is a diagram illustrating a learning phase process of an information processing device according to this embodiment. FIG. 2 is a diagram illustrating an analysis phase process of the information processing device according to this embodiment. FIG. 3 is a functional block diagram illustrating a configuration of an information processing device according to this embodiment. FIG. 4 is a diagram illustrating an example of a data structure of a word vector dictionary. FIG. 5A is a diagram (1) illustrating a process of calculating a sentence vector. FIG. 5B is a diagram (2) illustrating a process of calculating a sentence vector. FIG. 6 is a diagram illustrating a process of generating a sentence transposition index. FIG. 7 is a flowchart illustrating a processing procedure of a learning phase of an information processing device according to this embodiment. FIG. 8 is a flowchart illustrating a processing procedure of an analysis phase of an information processing device according to this embodiment. FIG. 9 is a diagram (1) illustrating other processing of the information processing device. FIG. 10 is a diagram (2) illustrating other processing of the information processing device. FIG. 11 is a diagram illustrating an example of a hardware configuration of a computer that realizes the same functions as the information processing device of the embodiment.

[0012] Hereinafter, an information processing program, an information processing method, and an information processing device disclosed in the present application will be described in detail with reference to the accompanying drawings. However, the present invention is not limited to these embodiments.

[0013] The process of the information processing device according to the present embodiment will be described. The information processing device executes the process of the learning phase and then executes the process of the analysis phase. Fig. 1 is a diagram for explaining the process of the learning phase of the information processing device according to the present embodiment.

[0014] In the learning phase, the information processing device performs learning of the machine learning model 50 (trains the machine learning model 50) using a plurality of sentences included in the teacher data 141. The machine learning model 50 is a neural network (NN) such as BERT (Pre-training of Deep Bidirectional Transformers for Language Understanding), Next Sentence Prediction, or Transformers.

[0015] The text included in the training data 141 includes multiple sentences. The multiple sentences have a predetermined relationship with the preceding and following sentences. Each sentence is a sentence that is set in advance based on an inductive or deductive syllogism, etc.

[0016] For example, sentence 10a includes, from the beginning, the sentences "Birds lay eggs," "Penguins are birds," ..., and "Therefore, penguins lay eggs." Sentence 10b includes, from the beginning, the sentences "Birds are born from eggs," "Pigeons are a type of bird," ..., and "Therefore, pigeons are born from eggs."

[0017] The information processing device calculates the sentence vector of each sentence included in the sentences 10 a, 10 b, and other sentences. For example, the information processing device performs morphological analysis on the sentences to divide them into words, and calculates the sentence vector by integrating the vectors of each word.

[0018] The sentence vector of the sentence 10a, "Birds lay eggs.", is set to "SV1-1." The sentence vector of the sentence "Penguins are birds." is set to "SV1-2." The sentence vector of the sentence "Therefore, penguins lay eggs." is set to "SV1-3."

[0019] The sentence vector of the sentence 10b "Birds hatch from eggs" is "SV2-1." The sentence vector of the sentence "Pigeons are a type of bird" is "SV2-2." The sentence vector of the sentence "Therefore, pigeons hatch from eggs" is "SV2-3."

[0020] The information processing device repeatedly executes a process of inputting the vectors of the sentences contained in a sentence to the machine learning model 50 in order, starting from the first sentence. For example, the information processing device inputs the sentence vectors to the machine learning model 50 in the order of sentence vectors "SV1-1", "SV1-2", ..., "SV1-3". The information processing device inputs the sentence vectors to the machine learning model 50 in the order of sentence vectors "SV2-1", "SV2-2", ..., "SV2-3".

[0021] By the information processing device executing the above-mentioned learning phase processing, a machine learning model 50 is generated that, when a sentence vector of a first sentence is input, predicts the sentence vector of the second sentence that follows the first sentence.

[0022] 2 is a diagram illustrating the processing of the analysis phase of the information processing device according to the present embodiment. In the analysis phase, the information processing device uses a trained machine learning model 50 to calculate sentence vectors included in the sentence to be processed and detects inappropriate sentences using cosine similarity or the like.

[0023] 2, the sentence to be processed that contains an input error or the like is referred to as sentence 20. Sentence 20 is composed of, from the beginning, the sentence "Birds lay eggs," the sentence "Penguins take pictures," ..., and the sentence "Therefore, penguins lay eggs." The sentence "Penguins take pictures" is a sentence that contains an input error of the homonym "picture" of the word "bird" compared to the correct sentence "Penguins are birds" included in sentence 10a of training data 141.

[0024] The information processing device calculates a sentence vector "SV1-1" of the sentence "Birds lay eggs." and inputs the calculated sentence vector "SV1-1" into the machine learning model 50 to predict the sentence vector of the sentence following the sentence "Birds lay eggs." In the example shown in FIG. 2 , the machine learning model 50 predicts "SV1-2" as the sentence vector of the sentence following the sentence "Birds lay eggs."

[0025] The information processing device calculates the sentence vector "SV3" of the sentence "Penguins are photographed," which is included in the text 20 and is the next sentence to the sentence "Birds lay eggs."

[0026] The information processing device calculates the cosine similarity between the sentence vector "SV1-2" of the next sentence predicted by the machine learning model 50 and the sentence vector "SV3" of the sentence "Penguins are photographed," which is included in the sentence 20 and is the next sentence after the sentence "Birds lay eggs."

[0027] The information processing device determines that the sentence "Penguins are birds," which is included in sentence 10a and follows the sentence "Birds lay eggs," is a correct (hereinafter referred to as "appropriate") sentence when the cosine similarity is less than the threshold. On the other hand, the information processing device determines that the sentence "Penguins are photographed," which is included in sentence 20 and follows the sentence "Birds lay eggs," is an inappropriate sentence containing an input error or the like when the cosine similarity is less than the threshold.

[0028] As described above, the information processing device inputs the vectors of each sentence of the sentences included in the training data 141 into the machine learning model 50 in order, thereby generating a machine learning model 50 that, when the sentence vector of a first sentence is input, predicts the sentence vector of the second sentence that follows the first sentence. The information processing device inputs the sentence vectors of the sentences of the processing target sentence into the generated machine learning model, predicts the sentence vector of the next sentence, and detects sentences with input errors from the processing target sentence based on the predicted sentence vector. In other words, it is possible to detect sentences that contain input errors or the like and have inappropriate sentence vectors from each sentence included in the processing target sentence.

[0029] In addition, if the information processing device determines in the processing of Figure 2 that the sentence "Penguins are photographs" is a sentence with an inappropriate sentence vector, it may search for a sentence with an appropriate sentence vector, "Penguins are birds," from a database (DB) or the like based on the sentence vector SV1-2 predicted by the machine learning model 50, and output it to the display device as a correct correction candidate (hereinafter referred to as "optimization").

[0030] Furthermore, the information processing device may calculate the word vectors of the multiple words "penguin," "ha," and "take." that make up the sentence "Penguin wa tora" (take a picture of a penguin) in which an inappropriate sentence vector was detected, using another machine learning model that has learned the order of word-based vectors, and correct input errors such as those for the deviating word "take."

[0031] Next, an example of the configuration of an information processing device that executes the processes described in Figures 1 and 2 will be described. Figure 3 is a functional block diagram showing the configuration of the information processing device according to this embodiment. As shown in Figure 3, the information processing device 100 includes a communication unit 110, an input unit 120, a display unit 130, a storage unit 140, and a control unit 150.

[0032] The communication unit 110 is connected to an external device or the like via a wired or wireless connection, and transmits and receives information to and from the external device or the like. For example, the communication unit 110 is realized by a network interface card (NIC) or the like. The communication unit 110 may be connected to a network (not shown).

[0033] The input unit 120 is an input device that inputs various types of information to the information processing device 100. The input unit 120 corresponds to a keyboard, a mouse, a touch panel, etc. For example, a user may operate the input unit 120 to input text data, etc.

[0034] The display unit 130 is a display device that displays information output from the control unit 150. The display unit 130 corresponds to a liquid crystal display, an organic EL (Electro Luminescence) display, a touch panel, etc. For example, a sentence containing an input error is displayed on the display unit 130.

[0035] The storage unit 140 includes a machine learning model 50, training data 141, and a word vector dictionary 142. The storage unit 140 is realized by, for example, a semiconductor memory element such as a random access memory (RAM) or a flash memory, or a storage device such as a hard disk or an optical disk.

[0036] The machine learning model 50 is a neural network such as BERT, Next Sentence Prediction, or Transformers, which are described in FIG. 1 .

[0037] The training data 141 is the training data 141 described in FIG. 1. The text included in the training data 141 includes multiple sentences. The multiple sentences have a predetermined relationship with the preceding and following sentences. Each sentence is a sentence that is set in advance based on an inductive or deductive syllogism, etc.

[0038] The word vector dictionary 142 is a table that defines the codes and word vectors assigned to words. FIG. 4 is a diagram showing an example of the data structure of the word vector dictionary. As shown in FIG. 4, the word vector dictionary 142 has codes, words, and word vectors (1) to (7). A code is a code assigned to a word. A word is a word included in a character string. Word vectors (1) to (7) are vectors assigned to words. The nth component of a word vector is denoted as word vector (n) (n = 1 to 7).

[0039] The DB 143 stores various texts. The texts include multiple sentences, and each sentence includes multiple words. The DB 143 may store the texts included in the training data 141.

[0040] The sentence transposition index 144 associates a sentence vector with a position pointer. The position pointer indicates the position in the DB 143 where the sentence corresponding to the sentence vector exists.

[0041] Returning to the description of Fig. 3, the control unit 150 includes a preprocessing unit 151, a learning unit 152, and an analysis unit 153. The control unit 150 is realized by, for example, a central processing unit (CPU) or a micro processing unit (MPU). The control unit 150 may also be implemented by an integrated circuit such as an application specific integrated circuit (ASIC) or a field programmable gate array (FPGA).

[0042] The preprocessing unit 151 executes various preprocessing operations. For example, the preprocessing unit 151 acquires unprocessed sentences from the DB 143 and calculates sentence vectors of the sentences. The preprocessing unit 151 sets the relationship between the calculated sentence vectors and the position pointers of the sentences corresponding to the sentence vectors in the sentence transposition index 144.

[0043] An example of processing by the preprocessing unit 151 to calculate a sentence vector of a sentence will be described. FIGS. 5A and 5B are diagrams for explaining the processing for calculating a sentence vector. Here, a case where the sentence vector of the sentence "Horses like carrots." is calculated will be described. The preprocessing unit 151 performs morphological analysis on the sentence "Horses like carrots." to break it down into multiple words. A "△ (space)" is added to each of the broken down words. For example, sentence 1 "Horses like carrots." is divided into "horse △", "is △", "carrot △", "ga △", "suki △", "desu △", and ". △".

[0044] The preprocessing unit 151 identifies a code corresponding to each word and replaces it with the word by comparing each divided word with the word vector dictionary 45. For example, the words "horse △", "is △", "carrot △", "ga △", "like △", "desu △", and ". △" are replaced with "C1", "C2", "C3", "C4", "C5", "C6", and "C7", respectively.

[0045] We now move on to the explanation of Figure 5B. The preprocessing unit 151 identifies word vectors (1) to (7) assigned to the codes based on the word vector dictionary 45 and each code. For example, the word vectors (1) to (7) for the code "C1" are designated wv1-1 to wv1-7. The word vectors (1) to (7) for the code "C2" are designated wv2-1 to wv2-7. The word vectors (1) to (7) for the code "C3" are designated wv3-1 to wv3-7.

[0046] The word vectors (1) to (7) of code "C4" are wv4-1 to wv4-7. The word vectors (1) to (7) of code "C5" are wv5-1 to wv5-7. The word vectors (1) to (7) of code "C6" are wv6-1 to wv6-7. The word vectors (1) to (7) of code "C7" are wv7-1 to wv7-7.

[0047] The preprocessing unit 151 calculates the sentence vector SV1 of the sentence by integrating the word vectors for each element. For example, the preprocessing unit 151 calculates the first component "SV1-1" of the sentence vector SV1 by integrating wv1-1 to wv7-1, which become the respective word vectors (1). The preprocessing unit 151 calculates the second component "SV1-2" of the sentence vector SV1 by integrating wv1-2 to wv7-2, which become the respective word vectors (2). The preprocessing unit 151 calculates the third component "SV1-3" of the sentence vector SV1 by integrating wv1-3 to wv7-3, which become the respective word vectors (3).

[0048] The preprocessing unit 151 calculates the fourth component "SV1-4" of the sentence vector SV1 by integrating wv1-4 to wv7-4, which become the respective word vectors (4). The preprocessing unit 151 calculates the fifth component "SV1-5" of the sentence vector SV1 by integrating wv1-5 to wv7-5, which become the respective word vectors (5). The preprocessing unit 151 calculates the sixth component "SV1-6" of the sentence vector SV1 by integrating wv1-6 to wv7-6, which become the respective word vectors (6). The preprocessing unit 151 calculates the seventh component "SV1-7" of the sentence vector SV1 by integrating wv1-7 to wv7-7, which become the respective word vectors (7).

[0049] The preprocessing unit 151 repeatedly executes the above process for each sentence in the other documents included in the DB 143, thereby calculating the sentence vector for each sentence.

[0050] The preprocessing unit 151 generates the sentence transposition index 144 by associating the calculated sentence vector of each sentence with a position pointer in the DB 143. The preprocessing unit 151 may generate the sentence transposition index 144 with a data structure as shown in FIG. 6. FIG. 6 is a diagram for explaining the process of generating the sentence transposition index. As shown in FIG. 6, the preprocessing unit 151 may associate the sentence vector with a plurality of record pointers and a plurality of position pointers, and associate each record pointer and position pointer with each sentence in the DB 143.

[0051] Returning to the explanation of Fig. 3, the learning unit 152 executes the processing of the learning phase explained in Fig. 1 to generate a machine learning model 50 that, when a sentence vector of a first sentence is input, predicts the sentence vector of a second sentence that follows the first sentence.

[0052] For example, the learning unit 152 calculates a sentence vector for each sentence included in the text of the teacher data 141 and inputs the calculated sentence vectors in order to the machine learning model 50, thereby performing learning of the machine learning model 50. Other processes of the learning unit 152 are similar to the processes described in Fig. 1. The process by which the learning unit 152 calculates the sentence vector of a sentence is similar to the process by which the preprocessing unit 151 calculates the sentence vector of a sentence.

[0053] The analysis unit 153 executes the processing of the analysis phase described with reference to FIG. 2 to detect sentences with inappropriate sentence vectors from among the sentences included in the sentence to be processed.

[0054] For example, when the analysis unit 153 receives the sentence 20 to be processed, it calculates the sentence vector of the sentence included in the sentence 20. The analysis unit 153 identifies the sentence included in the sentence 20 based on the period "." included in the sentence 20. The process by which the analysis unit 153 calculates the sentence vector of a sentence is similar to the process by which the preprocessing unit 151 calculates the sentence vector of a sentence. The sentence vector of the nth sentence from the beginning of the sentence 20 (n = 0 to M) is denoted as "SVn."

[0055] The analysis unit 153 inputs the sentence vector SVn into the trained machine learning model 50 to predict the sentence vector SVn+1' of the (n+1)th sentence from the beginning of the text 20. The analysis unit 153 calculates the cosine similarity between the sentence vector SVn+1' predicted using the machine learning model 50 and the sentence vector SVn+1.

[0056] If the cosine similarity between the sentence vector SVn+1' and the sentence vector SVn+1 is equal to or greater than a threshold, the analysis unit 153 determines that the n+1-th sentence from the top is an appropriate sentence. On the other hand, if the cosine similarity between the sentence vector SVn+1' and the sentence vector SVn+1 is equal to or greater than a threshold, the analysis unit 153 determines that the n+1-th sentence from the top is an inappropriate sentence.

[0057] If the analysis unit 153 determines that the sentence vector is an inappropriate sentence, it compares the sentence vector SVn+1' with the sentence transposition index 144 and identifies the position pointer of the sentence corresponding to the sentence vector SVn+1'. The analysis unit 153 searches the DB 143 for the sentence corresponding to the sentence vector SVn+1' based on the position pointer. The analysis unit 153 associates the inappropriate sentence of the sentence vector with the searched sentence and displays them on the display unit 130.

[0058] The analysis unit 153 may compare the inappropriate sentences in the sentence vector with the searched sentence on a word-by-word basis, detect incorrectly input words from the inappropriate sentences in the sentence vector, and display the detected words.

[0059] Next, an example of a processing procedure of the information processing device 100 according to this embodiment will be described. Fig. 7 is a flowchart showing the processing procedure of the learning phase of the information processing device according to this embodiment. As shown in Fig. 7, the learning unit 152 of the information processing device 100 selects an unselected sentence from the training data 141 (step S101).

[0060] The learning unit 152 calculates a sentence vector for each sentence included in the selected sentence, and generates a sentence transposition index that associates the sentence vector with a DB record and the position of the sentence (step S102).The learning unit 152 executes learning by inputting the sentence vectors of the selected sentences, starting from the first sentence, into the machine learning model 50 (step S103).

[0061] If the learning unit 152 determines to continue learning (Yes at step S104), the process proceeds to step S101. On the other hand, if the learning unit 152 determines not to continue learning (No at step S104), the process ends the learning phase.

[0062] 8 is a flowchart showing the processing procedure of the analysis phase of the information processing apparatus according to this embodiment, as shown in Fig. 8, the analysis unit 153 of the information processing apparatus 100 receives input of a sentence to be processed (step S201).

[0063] The analysis unit 153 calculates a sentence vector for each sentence included in the input text (step S202). The analysis unit 153 sets n to an initial value (step S203).

[0064] The analysis unit 153 inputs the sentence vector SVn of the nth sentence among the multiple sentences included in the document to the machine learning model 50, and predicts the sentence vector SVn+1' of the n+1th sentence (step S204).

[0065] The analysis unit 153 calculates the cosine similarity between the sentence vector SVn+1 of the (n+1)th sentence among the multiple sentences included in the document and the sentence vector SVn+1′ of the predicted sentence (step S205).

[0066] If the cosine similarity is equal to or greater than the threshold (Yes at Step S206), the analysis unit 153 proceeds to Step S210.

[0067] On the other hand, if the cosine similarity is not equal to or greater than the threshold (step S206, No), the analysis unit 153 detects the (n+1)th sentence as a sentence with an inappropriate sentence vector (step S207).The analysis unit 153 detects a sentence corresponding to the sentence vector SVn+1' from the DB 143 based on the predicted sentence vector SVn+1' and the sentence transposition index 144 (step S208).

[0068] The analysis unit 153 displays the sentence with the inappropriate sentence vector and the sentence detected from the DB 143 on the display unit 130 (step S209).

[0069] The processing from step S210 onwards will be described. If n is equal to or greater than L (step S210, Yes), the analysis unit 153 ends the processing. L is the number of sentences included in the text to be processed. If n is not equal to or greater than L (step S210, No), the analysis unit 153 updates n by adding 1 to the value n (step S211) and proceeds to step S204.

[0070] Next, the effects of the information processing device 100 according to this embodiment will be described. The information processing device 100 generates a machine learning model 50 that, when the sentence vector of a first sentence is input, predicts the sentence vector of the second sentence that follows the first sentence, by inputting the vector of each sentence in turn in the training data 141. The information processing device 100 inputs the sentence vector of the sentence to be processed into the generated machine learning model 50, predicts the sentence vector of the next sentence, and detects sentences with inappropriate sentence vectors from the sentence to be processed based on the predicted sentence vector. Furthermore, it is possible to correct words such as input errors from the inappropriate sentence.

[0071] The information processing device 100 detects an inappropriate sentence in the sentence vector and corrects input errors, etc., based on the cosine similarity between the sentence vector of the next sentence predicted by the machine learning model 50 and the sentence vector of the sentence following the sentence included in the sentence to be processed. This makes it possible to detect an inappropriate sentence in the sentence vector and correct input errors, etc., while suppressing calculation costs.

[0072] The information processing device 100 inputs vectors of a plurality of sentences, the order of which has been determined based on induction or deduction, into a machine learning model in order to train it, thereby making it possible to predict the next sentence of a target sentence based on induction or deduction.

[0073] When the information processing device 100 determines that the sentence is to be corrected, it searches for the corrected sentence based on the vector predicted by the machine learning model 50. This makes it possible to notify the corrected sentence.

[0074] The above-described processing contents of the information processing device 100 are merely examples, and the information processing device 100 may execute other processes. The other processes of the information processing device 100 will be described below.

[0075] 9 and 10 are diagrams for explaining other processing of the information processing device. The information processing device 100 described above has the machine learning model 50 learn the order of vectors of sentences based on a syllogism. However, instead of sentence vectors, the machine learning model 50 may learn the order of vectors of protein primary structures, which are protein sequences and comprise multiple amino acid sequences corresponding to words. In the following description, the consecutive amino acid sequence of a protein will be referred to as a "basic structure," and the protein primary structure will be referred to as a "primary structure."

[0076] 9 will be described. In the learning phase, the information processing device 100 executes learning of the machine learning model 50 using a plurality of protein sequences 20 a and 20 b included in the training data 241.

[0077] For example, sequence 20a includes the primary structures "α primary structure," "β primary structure," ..., "γ primary structure." Sequence 20b includes the primary structures "Δ primary structure," "ε primary structure," ..., "ζ primary structure."

[0078] The information processing device 100 identifies the vector of each primary structure using a vector dictionary of protein basic structures that associates basic structures with vectors. For example, the vector of a primary structure "α primary structure" composed of multiple basic structures is "V20-1", the vector of a primary structure "β primary structure" is "V20-2", and the vector of a primary structure "γ primary structure" is "V20-3". The vector of a primary structure "Δ primary structure" is "V21-1", the vector of a primary structure "ε primary structure" is "V21-2", and the vector of a primary structure "ζ primary structure" is "V21-3". The vector of each primary structure is calculated based on the vectors of each of the multiple basic structures that make up that primary structure.

[0079] The information processing device 100 repeatedly executes a process of inputting vectors of the primary structure included in the protein sequence into the machine learning model 50 in order, starting from the first vector. For example, the information processing device inputs vectors into the machine learning model 50 in the order of vectors "V20-1", "V20-2", ..., "V20-3". The information processing device inputs vectors into the machine learning model 50 in the order of vectors "V21-1", "V21-2", ..., "V21-3".

[0080] By the information processing device 100 executing the above-mentioned learning phase processing, a machine learning model 50 is generated that predicts the vector of the next primary structure of a given primary structure when a vector of a given primary structure is input.

[0081] 10 will be described. In the analysis phase, the sequence of the protein to be processed is defined as sequence 25. Sequence 25 includes, from the beginning, the primary structures "α primary structure", "η primary structure", ..., "γ primary structure".

[0082] The information processing device 100 calculates a vector "V20-1" of the primary structure "α primary structure" and inputs the calculated vector "V20-1" into the machine learning model 50 to predict the vector of the primary structure next to the primary structure "α primary structure". In the example shown in FIG. 10 , the machine learning model 50 predicts "V20-2" as the vector of the primary structure next to the primary structure "α primary structure".

[0083] The information processing device 100 calculates the vector "V22" of the "η primary structure" which is a primary structure included in the sequence 25 and which is next to the primary structure "α primary structure".

[0084] The information processing device 100 calculates the cosine similarity between the vector "V20-2" of the next primary structure predicted by the machine learning model 50 and the vector "V22" of the "η primary structure" which is the primary structure included in the sequence 25 and is the next of the basic structure "α primary structure."

[0085] When the cosine similarity is equal to or greater than the threshold, the information processing device determines that the "η primary structure" next to the primary structure "α primary structure" contained in the sequence 25 is the correct primary structure. On the other hand, when the cosine similarity is less than the threshold, the information processing device determines that the "η primary structure" next to the primary structure "α primary structure" contained in the sequence 25 is an incorrect primary structure, and corrects mutations or the like in the basic structure contained in the primary structure "η primary structure."

[0086] By executing the processes shown in Figures 9 and 10 with the information processing device 100, primary structures with inappropriate primary structure vectors can be detected from multiple primary structures contained in a protein sequence, and basic structures with mutations, etc. can be optimized. This makes it possible to detect protein primary structures with mutations, etc. (SNPs are a typical example) that have occurred in receptors composed of multiple primary protein structures. Furthermore, by machine learning the binding order of the multiple primary protein structures that make up the receptor and the single or multiple primary protein structures that bind to the receptor, it is possible to predict the protein primary structure vector of a ligand that binds to the receptor. This can support the improvement of ligands with new primary protein structure vectors that are similar to ligands already commercialized as biopharmaceuticals, have excellent medicinal efficacy, and suppress side effects.

[0087] Next, an example of the hardware configuration of a computer that realizes the same functions as the information processing device 100 described in the above embodiment will be described. Fig. 11 is a diagram showing an example of the hardware configuration of a computer that realizes the same functions as the information processing device of the embodiment.

[0088] 11, computer 300 has a CPU 301 that executes various types of arithmetic processing, an input device 302 that accepts data input from a user, and a display 303. Computer 300 also has a communication device 304 that exchanges data with external devices via a wired or wireless network, and an interface device 305. Computer 300 also has a RAM 306 that temporarily stores various types of information, and a hard disk drive 307. Each of devices 301 to 307 is connected to a bus 308.

[0089] The hard disk drive 307 stores a preprocessing program 307 a, a learning program 307 b, and an analysis program 307 c. The CPU 301 reads out each of the programs 307 a to 307 c and loads them into the RAM 306.

[0090] The pre-processing program 307a functions as the pre-processing process 306a, the learning program 307b functions as the learning process 306b, and the analysis program 307c functions as the analysis process 306c.

[0091] The processing of the preprocessing process 306a corresponds to the processing of the preprocessing unit 151. The processing of the learning process 306b corresponds to the processing of the learning unit 152. The processing of the analysis process 306c corresponds to the processing of the analysis unit 153.

[0092] It should be noted that each of the programs 307a to 307c does not necessarily have to be stored in the hard disk drive 307 from the beginning. For example, each program may be stored in a "portable physical medium" such as a flexible disk (FD), CD-ROM, DVD, magneto-optical disk, or IC card that is inserted into the computer 300. Then, the computer 300 may read and execute each of the programs 307a to 307c.

[0093] 50 Machine learning model 100 Information processing device 110 Communication unit 120 Input unit 130 Display unit 140 Storage unit 141 Teacher data 142 Word vector dictionary 143 DB 144 Sentence transposition index 150 Control unit 151 Preprocessing unit 152 Learning unit 153 Analysis unit

Claims

1. Calculating vectors of a plurality of consecutive sentences, each of which has a relationship with the preceding and following sentences; inputting the vectors of the plurality of sentences into a machine learning model in order to train the model, thereby generating the machine learning model that, when a vector of a sentence is input into the machine learning model, predicts the sentence vector of a sentence that is input next to the given sentence; calculating a vector of a first sentence and a vector of a second sentence following the first sentence; The vector of the first sentence is input to the machine learning model to calculate a vector of a sentence predicted to follow the first sentence, and it is determined whether the vector of the second sentence is correct. An information processing program that causes a computer to execute a process.

2. The information processing program according to claim 1, characterized in that the determining process determines whether the vector of the second sentence is appropriate based on the cosine similarity between a vector predicted by inputting the vector of the first sentence into the machine learning model and the vector of the second sentence.

3. The information processing program according to claim 2, characterized in that the consecutive sentences are sentences whose order is determined based on an inductive or deductive method, and the process of generating the machine learning model includes inputting vectors of the sentences whose order is determined based on an inductive or deductive method in order into the machine learning model for training.

4. The information processing program according to claim 1, further comprising: inputting the vector of the first sentence into the machine learning model to calculate a vector of a sentence predicted to follow the first sentence when the vector of the second sentence is determined to be inappropriate; searching for sentences similar to the calculated vector and presenting them as candidates for appropriate sentences; and causing the computer to execute a process of recommending an appropriate sentence based on the calculated vector of the sentence predicted to follow the first sentence.

5. Calculating vectors of a plurality of consecutive sentences, each of which has a relationship with the preceding and following sentences; inputting the vectors of the plurality of sentences into a machine learning model in order to train the model, thereby generating the machine learning model that, when a vector of a sentence is input into the machine learning model, predicts the sentence vector of a sentence that is input next to the given sentence; calculating a vector of a first sentence and a vector of a second sentence following the first sentence; The vector of the first sentence is input to the machine learning model to calculate a vector of a sentence predicted to follow the first sentence, and it is determined whether the vector of the second sentence is correct. An information processing method characterized in that the processing is executed by a computer.

6. Calculating vectors of a plurality of consecutive sentences, each of which has a relationship with the preceding and following sentences; inputting the vectors of the plurality of sentences into a machine learning model in order to train the model, thereby generating the machine learning model that, when a vector of a sentence is input into the machine learning model, predicts the sentence vector of a sentence that is input next to the given sentence; calculating a vector of a first sentence and a vector of a second sentence following the first sentence; The vector of the first sentence is input to the machine learning model to calculate a vector of a sentence predicted to follow the first sentence, and it is determined whether the vector of the second sentence is correct.

1. An information processing device comprising: a control unit that executes processing.