Information processing device, information processing method, and program

Abstract

In order to solve the problem of providing a technique that has an excellent balance of computational cost and processing power and that is applicable to natural language processing in a medical setting, the information processing device of the present invention comprises: an acquisition means (21) for acquiring a token sequence obtained from a medical sentence in an electronic medical record, and a context information vector obtained from context information of the electronic medical record; and an output sequence generation means (22) for executing output sequence generation processing for generating an output sequence from the token sequence and the context information vector, the output sequence generation processing including processing for converting to a high-dimension characteristic vector having a higher dimension than the sum of the dimension of the token sequence and the dimension of the context information vector.

Description

[Technical Field]

[0001] The present invention relates to an information processing device, an information processing method, and a program that have the processing power to grasp the meaning of a sentence more accurately and realize language processing with an excellent balance between calculation cost and processing power. [Background technology]

[0002] In recent years, natural language processing technology using deep learning has been put into practical use, and as a result of its significant improvement in capabilities, it is being applied to various fields. For example, the application of natural language processing using deep learning is also being promoted in medical documents.

[0003] For example, a technology has been proposed that uses a similarity index value calculated by analyzing the text contained in a patient's electronic medical record to predict the possibility that the patient will engage in risky behavior based on the content of the text contained in the electronic medical record (see, for example, Patent Document 1).

[0004] In addition, a technology has been proposed that calculates the similarity between medical procedures recorded in medical data (e.g., prescription information, electronic medical record information) based on other medical procedures and names of illnesses or injuries recorded around the medical procedure, and extracts similar medical procedures as a group (see, for example, Patent Document 2). [Prior art documents] [Patent documents]

[0005] [Patent Document 1] International Patent Application No. WO2019 / 212005 [Patent Document 2] Japanese Patent Application No. 2019-212034 Summary of the Invention [Problem to be solved by the invention]

[0006] However, for example, in order to accurately analyze medical texts in Patent Document 1, an extremely high-performance information processing device such as a GPU is required, which poses a problem of a large cost burden on medical sites. Also, the technology in Patent Document 2 does not take into account the order in which words appear in medical texts, so there is a possibility that medical texts with different meanings may be judged to be similar medical texts.

[0007] One aspect of the present invention has been made in consideration of the above-mentioned problems, and one of its objectives is to provide a technology that has an excellent balance between computational cost and processing power and is applicable to natural language processing in medical settings. [Means for solving the problem]

[0008] An information processing device according to one aspect of the present invention comprises an acquisition means for acquiring a token sequence obtained from medical text in an electronic medical record and a context information vector obtained from context information in the electronic medical record, and an output sequence generation means for executing an output sequence generation process for generating an output sequence from the token sequence and the context information vector, the output sequence generation process including a process for converting the token sequence into a high-dimensional feature vector having a dimension higher than the sum of the dimensions of the token sequence and the context information vector.

[0009] An information processing method according to one aspect of the present invention includes acquiring a token sequence obtained from medical text in an electronic medical record and a context information vector obtained from context information in the electronic medical record, and performing an output sequence generation process to generate an output sequence from the token sequence and the context information vector, the output sequence generation process including a process of converting the token sequence into a high-dimensional feature vector having a dimension higher than the sum of the dimensions of the token sequence and the dimensions of the context information vector.

[0010] A program according to one aspect of the present invention causes a computer to function as an information processing device that includes: an acquisition means for acquiring a token sequence obtained from medical text in an electronic medical record and a context information vector obtained from context information in the electronic medical record; and an output sequence generation means for executing an output sequence generation process that generates an output sequence from the token sequence and the context information vector, the output sequence generation process including a process of converting the token sequence into a high-dimensional feature vector having a dimension higher than the sum of the dimensions of the token sequence and the context information vector. [Effects of the Invention]

[0011] According to one aspect of the present invention, it is possible to provide a technology that has an excellent balance between calculation cost and processing power and is applicable to natural language processing in medical settings. [Brief explanation of the drawings]

[0012] [Figure 1] 1 is a block diagram showing an example of the configuration of an information processing device according to a first exemplary embodiment of the present invention. [Figure 2] 1 is a flowchart showing the flow of an information processing method according to the first exemplary embodiment of the present invention. [Figure 3] FIG. 10 is a block diagram showing an example of the configuration of a natural language processing apparatus according to a second exemplary embodiment of the present invention. [Figure 4] FIG. 4 is a diagram modeling the output sequence generation process executed by the output sequence generator. [Figure 5] 10 is a flowchart illustrating an example of natural language processing according to the second exemplary embodiment of the present invention. [Figure 6] 10 is a flowchart illustrating an example of an output process. [Figure 7] 10 is a flowchart illustrating an example of a high-dimensional feature vector conversion process. [Figure 8] 1 is a diagram illustrating an example of data input to a natural language processing device and data output as a result of executing natural language processing. FIG. [Figure 9] FIG. 1 is a diagram illustrating an input vector space and a high-dimensional feature space. [Figure 10] FIG. 10 is a block diagram showing an example of the configuration of a natural language learning and processing device according to a third exemplary embodiment of the present invention. [Figure 11] 10 is a flowchart illustrating an example of a learning process according to the third exemplary embodiment of the present invention. [Figure 12] 10 is a flowchart illustrating an example of a parameter update process. [Figure 13] 10 is a flowchart illustrating an example of an output weight matrix learning process. [Figure 14] FIG. 10 is a diagram showing an example of medical text in which the same words are used. [Figure 15] FIG. 10 is a block diagram showing an example of the configuration of a natural language learning device according to an exemplary embodiment 4 of the present invention. [Figure 16] FIG. 2 is a block diagram showing an example of a hardware configuration of an apparatus in each exemplary embodiment of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0013] Exemplary Embodiment 1 A first exemplary embodiment of the present invention will be described in detail with reference to the drawings. This exemplary embodiment is a basic form of the exemplary embodiments described below.

[0014] <Overview of information processing device 20> The information processing device 20 according to this exemplary embodiment is, in broad terms, a device that predicts whether the words that make up a given sentence are used in the sentence in a positive or negative sense.

[0015] More specifically, the information processing device 20, for example, an acquisition means for acquiring a token sequence obtained from the medical text of the electronic medical record and a context information vector obtained from the context information of the electronic medical record; and output sequence generation means for executing an output sequence generation process that generates an output sequence from a token sequence and a context information vector, the output sequence generation process including a process of converting the token sequence into a high-dimensional feature vector having a dimension higher than the sum of the dimensions of the token sequence and the dimensions of the context information vector.

[0016] <Configuration of information processing device 20> The configuration of an information processing device 20 according to this exemplary embodiment will be described with reference to Fig. 1. Fig. 1 is a block diagram showing an example of the configuration of the information processing device 20.

[0017] 1, information processing device 20 includes an acquisition unit 21 and an output sequence generation unit 22. Acquisition unit 21 is configured to realize acquisition means in this exemplary embodiment. Output sequence generation unit 22 is configured to realize output sequence generation means in this exemplary embodiment.

[0018] The acquisition unit 21 acquires a token sequence obtained from the medical text of the electronic medical record and a context information vector obtained from the context information of the electronic medical record.

[0019] Here, an electronic medical record is, for example, data in which medical information including medical documents is digitized, structured, and recorded.

[0020] A token string is a sequence of multiple tokens. Tokens are generated, for example, by vectorizing each word included in a medical text, and then arranging these tokens in the order in which each word appears to generate a token string.

[0021] Context information is, for example, structured information contained in a structured electronic medical record, such as information other than the text recorded as medical text. For example, context information includes information such as the job title of the person who wrote the medical text and the date and time the medical text was recorded. Context information is also vectorized into a context information vector.

[0022] The output sequence generation unit 22 executes an output sequence generation process that generates an output sequence from a token sequence and a context information vector, and that process includes a process of converting the token sequence into a high-dimensional feature vector having a dimension higher than the sum of the dimensions of the token sequence and the context information vector.

[0023] The high-dimensional feature vector is generated by, for example, multiplying the token and vectorized context information by a predetermined weight matrix, and the output sequence is generated by multiplying the high-dimensional feature vector by a weight matrix obtained by prior learning. At this time, the weight matrix is ​​determined so that the dimension of the high-dimensional feature vector is higher than the sum of the dimensions of the vectors that make up each token in the token sequence and the dimension of the context information vector.

[0024] <Effects of the information processing device 20> According to the information processing device 20 of this exemplary embodiment, an output sequence generation process is executed to generate an output sequence from a token sequence and a context information vector, and the output sequence generation process includes a process of converting the token sequence into a high-dimensional feature vector having a dimension higher than the sum of the dimensions of the token sequence and the context information vector, thereby making it possible to grasp the meaning of medical text more accurately and easily.

[0025] In other words, an output sequence generation process is performed that includes a process of converting the token sequence into a high-dimensional feature vector having a dimension higher than the sum of the dimensions of the token sequence and the context information vector, thereby enabling a more accurate understanding of the meaning of the sentence. Furthermore, in this embodiment, an output sequence is generated from the token sequence and the context information vector. This makes it possible to more easily understand the meaning of the medical sentence compared to, for example, understanding the meaning of the medical sentence from only the words in the medical sentence.

[0026] <Flow of information processing method by information processing device 20> The flow of the information processing method executed by the information processing device 20 configured as above will be described with reference to Fig. 2. Fig. 2 is a flowchart showing the flow of the information processing method. As shown in the figure, the information processing includes steps S1 and S2.

[0027] In step S1, the acquisition unit 21 acquires a token sequence obtained from the medical text of the electronic medical record and a context information vector obtained from the context information of the electronic medical record.

[0028] In step S2, the output sequence generation unit 22 executes an output sequence generation process that generates an output sequence from a token sequence and a context information vector, the output sequence generation process including a process of converting the token sequence into a high-dimensional feature vector having a dimension higher than the sum of the dimensions of the token sequence and the context information vector.

[0029] <Effects of information processing methods> According to the information processing method of this exemplary embodiment, an output sequence generation process is performed that includes a process of converting a token sequence into a high-dimensional feature vector having a dimension higher than the sum of the dimensions of the token sequence and the context information vector, thereby enabling a more accurate understanding of the meaning of the sentence. Furthermore, in this embodiment, the output sequence is generated from the token sequence and the context information vector. This makes it easier to understand the meaning of the medical sentence compared to, for example, understanding the meaning of the medical sentence from only the words in the medical sentence.

[0030] Exemplary Embodiment 2 A second exemplary embodiment of the present invention will be described in detail with reference to the drawings. Note that components having the same functions as those described in the first exemplary embodiment are given the same reference numerals, and their description will be omitted as appropriate.

[0031] <Configuration of the natural language processing device 10> The configuration of a natural language processing apparatus 10 according to this exemplary embodiment will be described with reference to Fig. 3. Fig. 3 is a block diagram showing an example configuration of the natural language processing apparatus 10. As shown in Fig. 3, the natural language processing apparatus 10 includes an information processing apparatus 20, a storage unit 30, a communication unit 41, an input unit 42, and an output unit 43.

[0032] The information processing device 20 is a functional block having the same functions as the information processing device 20 described in the first exemplary embodiment.

[0033] The storage unit 30 is configured, for example, with a semiconductor memory device or the like, and stores data. In this example, the storage unit 30 stores electronic medical record data and model parameters. Here, the model parameters are weighting coefficients obtained by machine learning, which will be described later.

[0034] The communication unit 41 is an interface for connecting the natural language processing apparatus 10 to a network. The specific configuration of the network does not limit the present exemplary embodiment, but as an example, a wireless LAN (Local Area Network), a wired LAN, a WAN (Wide Area Network), a public line network, a mobile data communication network, or a combination of these networks can be used.

[0035] The input unit 42 accepts various inputs to the natural language processing apparatus 10. The specific configuration of the input unit 42 is not limited to this exemplary embodiment, but as an example, the input unit 42 may be configured to include input devices such as a keyboard and a touchpad. The input unit 42 may also be configured to include a data scanner that reads data via electromagnetic waves such as infrared rays or radio waves, and a sensor that senses the environmental state.

[0036] The output unit 43 is a functional block that outputs the processing results of the natural language processing device 10. The specific configuration of the output unit 43 is not limited to this exemplary embodiment, but as an example, the output unit 43 is configured by a display, a speaker, a printer, etc., and displays various processing results of the natural language processing device 10 on a screen or outputs them as voice or figures.

[0037] In the example of FIG. 3, the acquisition unit 21 includes a token sequence generation unit 61 and a context information vector generation unit 62.

[0038] The token sequence generator 61 generates a token sequence by converting each word contained in the medical text into a token by embedding it in a predefined first vector space. Here, the first vector space is configured so that words that may be contained in the medical text can be embedded. As an example, the first vector space is obtained by learning words that may be contained in the medical text in advance using a language processing model.

[0039] Although this language processing model does not limit this exemplary embodiment, Skipgram can be used as an example. Skipgram is a type of Word2vec, and is a language processing model that can predict what words are likely to appear around a given input word. Note that the number of dimensions of the vectors in the first vector space is arbitrary, but in order to perform the calculation processing described below at high speed, it is desirable to avoid an excessively high dimensionality.

[0040] The context information vector generation unit 62 extracts predetermined context information from the electronic medical record and generates a context information vector by embedding the context information in a predefined second vector space. As described above, the context information is, for example, information other than the text recorded as the medical text, and is information indicating the attributes of the medical text. For example, if the context information is the job title of the person who wrote the medical text, information such as "doctor," "nurse," "paramedic," "physiotherapist," "nutritionist," etc. is embedded in the second vector space to generate a context information vector.

[0041] The second vector space may be obtained by prior learning or may be a preset vector space. The number of dimensions of the vectors in the second vector space is arbitrary, but it is desirable to avoid an excessively high number of dimensions in order to perform the calculation processing described later at high speed.

[0042] In this way, the words and context information can be properly vectorized.

[0043] 3, the output sequence generation unit 22 included in the information processing device 20 includes an input weight matrix multiplication unit 81, a connection weight matrix multiplication unit 82, and an output weight matrix multiplication unit 83. The output sequence generation unit predicts (generates) an output sequence using, for example, an Echo State Network (ESN), which is one model of an RNN (Recurrent Neural Network).

[0044] An ESN consists of an input layer, a reservoir layer, and an output layer. It enables linear separation of input information by mapping the low-dimensional vectors in the input layer to a high-dimensional neuron state space using a nonlinear transformation. In addition, when learning the correlation between input and output data, only the output layer needs to be trained; the input and reservoir layers do not. This reduces the computational cost of machine learning compared to conventional RNNs.

[0045] The input weight matrix multiplication unit 81 calculates the product of a predetermined input weight matrix and a token, and calculates the product of a predetermined context weight matrix and a context information vector. Note that the input weight matrix and the context weight matrix can be set arbitrarily as matrices that do not change the number of dimensions of the token and the context information vector, respectively. In other words, there is no need to learn the input weight vector and the context weight matrix.

[0046] The product vectors thus obtained are combined to generate the input vector.

[0047] The connection weight matrix multiplication unit 82 converts the input vector into a high-dimensional feature vector by multiplying the input vector by a predetermined connection weight matrix. As described above, the high-dimensional feature vector has a dimension higher than the sum of the dimension of the token and the dimension of the context information vector. For example, if the dimension of the token as a vector is 3 and the dimension of the context information vector is 1, the high-dimensional feature vector will have a dimension higher than 4 (1 + 3), i.e., a vector with 5 or more dimensions.

[0048] In this way, tokens and context information can be transformed into high-dimensional vectors.

[0049] The connection weight matrix can be arbitrarily set as a matrix that transforms the dimension of the input vector into the dimension of the high-dimensional feature vector, i.e., there is no need to learn the connection weight matrix.

[0050] The output weight matrix multiplication unit 83 multiplies the high-dimensional feature vector by an output weight matrix obtained by prior learning. As a result, the high-dimensional feature vector is converted into a lower-dimensional output vector, and an output sequence is generated by arranging each output vector in accordance with the order of the token sequence. The output weight matrix is ​​a matrix that converts the dimension of the high-dimensional feature vector into the dimension of the output vector. The components of the output weight matrix are obtained by prior learning and stored as model parameters in the storage unit 30 in FIG. 3.

[0051] In this way, an output sequence is generated by multiplying a high-dimensional feature vector by an output weight matrix obtained through prior learning, so the only weight matrix that needs to be learned is the output weight matrix.

[0052] 4 is a diagram modeling the output sequence generation process executed by output sequence generator 22. The model shown in FIG. 4 uses an ESN and has an input layer 111, a reservoir layer 112, and an output layer 113.

[0053] In this example, each token in the input token sequence is assumed to be a three-dimensional vector, and is represented by input u(n). Here, "n" is a value indicating the order of the token in the token sequence, and represents, for example, the order in which the word corresponding to that token appears in a medical document. In the input layer 111, the input u(n) is multiplied by an input weight matrix Win. The input weight matrix Win is a matrix that does not change the dimensions of the token as a vector, and is set arbitrarily.

[0054] In this example, the context information vector is assumed to be a one-dimensional vector, and is indicated by context vector c. In the input layer 111, the context vector c is multiplied by a context weight matrix Wc. The context weight matrix Wc is a matrix that does not change the dimension of the context information vector, and is set arbitrarily.

[0055] The input vector is a vector obtained by combining a three-dimensional vector obtained by multiplying the input u(n) by the input weight matrix Win and a one-dimensional vector obtained by multiplying the context vector c by the context weight matrix Wc. In this example, the input vector is a four-dimensional vector.

[0056] In the reservoir layer 112, the input vector is multiplied by a connection weight matrix W to convert it into a high-dimensional feature vector. In this example, the input vector is converted into a seven-dimensional high-dimensional feature vector. The connection weight matrix W is a matrix that converts the dimensions of the input vector (four dimensions) into the dimensions of the high-dimensional feature vector (seven dimensions), and is set arbitrarily.

[0057] In the output layer 113, the high-dimensional feature vector is multiplied by an output weight matrix Wout to generate an output vector. Here, the output vector is assumed to be a three-dimensional vector and is represented by output y(n). The output weight matrix Wout is a matrix that converts the dimensions (7 dimensions) of the high-dimensional feature vector into the dimensions (3 dimensions) of the output vector, and the elements of the matrix are set by machine learning performed in advance. The machine learning of the output weight matrix is ​​performed, for example, by referring to training data including multiple pairs of medical text, context information, and positive or negative labels for predetermined words contained in the medical text. Details of this machine learning will be described later in the exemplary embodiment below.

[0058] Here, if the state vector of the element in the reservoir layer 112 is represented by x(n), the state equation of the reservoir layer 112 is expressed by the following equation.

[0059]

number

[0060] <Flow of natural language processing by the natural language processing device 10> The flow of natural language processing executed by the natural language processing apparatus 10 configured as above will be described with reference to FIGS.

[0061] The natural language processing shown in FIG. 5 is similar to the processing described above with reference to FIG. 2, and therefore a detailed description thereof will be omitted.

[0062] Fig. 6 is a flowchart illustrating an example of the output column generation process executed in step S12 of Fig. 5. In the example of Fig. 6, the output column generation process includes steps S31 to S37.

[0063] In step S31, the token sequence generation unit 61 extracts words from the medical text in the electronic medical record acquired in step S11 of Fig. 5. At this time, the token sequence generation unit 61 extracts words contained in the medical text by, for example, performing morphological analysis on the medical text recorded as text data. Note that in step S31, it is assumed that, for example, n words contained in the medical text are extracted.

[0064] In step S32, the token sequence generator 61 generates a token sequence by converting each word contained in the medical text into a token by embedding it in a predefined first vector space. As described above, the first vector space is obtained by previously learning words that may be contained in the medical text using a language processing model. This language processing model is not intended to limit the present exemplary embodiment, but Skipgram can be used as an example.

[0065] In step S33, the context information vector generation unit 62 extracts predetermined context information from the electronic medical record.

[0066] In step S34, the context information vector generation unit 62 generates a context information vector by embedding the context information in a predefined second vector space.

[0067] In this way, the words and context information are properly vectorized.

[0068] In step S35, the input weight matrix multiplication unit 81 and the connection weight matrix multiplication unit 82 execute high-dimensional feature vector conversion processing. Here, an example of the high-dimensional feature vector conversion processing in step S35 of Fig. 6 will be described with reference to the flowchart of Fig. 7.

[0069] In step S51, the input weight matrix multiplication unit 81 multiplies the tokens in the token sequence by a predetermined input weight matrix. This process corresponds to the process in which the input layer 111 shown in FIG. 4 multiplies the input u(n) by the input weight matrix Win.

[0070] In step S52, the input weight matrix multiplication unit 81 multiplies a predetermined context information vector by a context weight matrix. This process corresponds to the process of multiplying the context vector c by the context weight matrix Wc in the input layer 111 shown in FIG. 4.

[0071] In step S53, the input weight matrix multiplication unit 81 multiplies the input vector, which includes the product of the input weight matrix and the token and the product of the context weight matrix and the context information vector, by a predetermined connection weight matrix. This process corresponds to the process in which the input vector is multiplied by the connection weight matrix W in the reservoir layer 112 shown in FIG. 4.

[0072] In step S54, the input weight matrix multiplication unit 81 generates a high-dimensional feature vector.

[0073] In this way, a high-dimensional feature vector conversion process is performed, whereby the tokens and context information are converted into a high-dimensional vector.

[0074] Returning to FIG. 6, in step S36, the output weight matrix multiplication unit 83 multiplies the high-dimensional feature vector obtained as a result of the processing in step S35 by the output weight matrix obtained by prior learning. The output weight matrix multiplication unit 83 reads out the model parameters stored in the storage unit 30 and uses them as components of the output weight matrix. This processing corresponds to the processing in the output layer 113 shown in FIG. 4, in which the high-dimensional feature vector is multiplied by the output weight matrix Wout. As a result, the output vector y(n) is generated.

[0075] In this way, an output sequence is generated by multiplying a high-dimensional feature vector by an output weight matrix obtained through prior learning, so the only weight matrix that needs to be learned is the output weight matrix.

[0076] In step S37, the output weight matrix multiplication unit 83 generates an output sequence by arranging the output vectors in accordance with the order of the token sequence.

[0077] The output sequence does not necessarily have to include output vectors corresponding to all tokens included in the token sequence. For example, the output sequence may include only output vectors corresponding to nouns among the words in the sentence.

[0078] In this way, the output column generation process is performed.

[0079] 8 is a diagram illustrating an example of data input to the natural language processing device 10 and data output as a result of natural language processing performed by the natural language processing device 10. In this example, text data 151, which is a medical document from an electronic medical record and states, "No cough, tremors, or chills, but cloudy urine.", is input to the natural language processing device 10. The output data is, for example, an output string 152, which is data such as "cough -, tremors -, chills -, cloudy urine +."

[0080] In this output string, "-" represents a negative label, and "+" represents a positive label. In other words, in the input medical text, the words "cough," "tremors," and "chills" are each used with a negative meaning, while the word "cloudy urine" is used with a positive meaning. In other words, in this example, the symptoms or conditions of the patient corresponding to the electronic medical record do not include "cough," "tremors," or "chills," but do include "cloudy urine."

[0081] Each output vector included in the output sequence is a vector that represents a combination of a word and a label.

[0082] 4, in this embodiment, the input vectors in the input layer are converted into high-dimensional feature vectors in the reservoir layer 112. In this way, it becomes easier to classify the features of words that were difficult to discern in the input layer 111 based on some criteria in the reservoir layer 112.

[0083] For example, as shown in FIG. 9, consider an input vector space (e.g., the input layer 111 in FIG. 4) which is a low-dimensional space, and a high-dimensional feature space (e.g., the reservoir layer 112 in FIG. 4). In the example of FIG. 9, the input vector space is shown in a two-dimensional space, and the high-dimensional feature space is shown in a three-dimensional space. Furthermore, the black and white circles in the figure represent words in a sentence. In the input vector space, each word is scattered in a two-dimensional space, making it difficult to distinguish the features of each word.

[0084] By converting the input vector space into a high-dimensional feature space in the reservoir layer, it becomes easier to classify the features of each word based on some criteria. In the example in Figure 8, the features of words represented by black circles and words represented by white circles are classified according to the planes in the high-dimensional feature space in the three-dimensional space.

[0085] Furthermore, the output weight matrix multiplication unit 83 may further add, average, combine, or the like the output vectors generated in the processing of step S36 to generate a sentence vector corresponding to the input medical sentence. For example, by comparing the sentence vectors of two medical sentences using the information processing device 20, it is possible to determine the semantic similarity between those sentences.

[0086] <Natural language processing devices and the effects of natural language processing> As described above, the natural language processing device and natural language processing according to this exemplary embodiment employs an RNN language processing model consisting of an input layer, a reservoir layer, and an output layer. This allows for more accurate analysis of the meaning of sentences compared to language processing models such as Word2vec, while significantly reducing the computational cost during training.

[0087] In addition, the natural language processing apparatus and the natural language processing according to the present exemplary embodiment employ a configuration in which an input vector consisting of a token and a context information vector is converted into a high-dimensional feature vector, which allows for more accurate prediction of the label of the word compared to when only the words in a medical document are input.

[0088] Exemplary Embodiment 3 Next, a third exemplary embodiment of the present invention will be described in detail with reference to the drawings. Note that components having the same functions as those described in the first and second exemplary embodiments are denoted by the same reference numerals, and their description will be omitted as appropriate.

[0089] <Configuration of the natural language learning and processing device 10A> The configuration of a natural language learning and processing device 10A according to this exemplary embodiment will be described with reference to Fig. 10. The natural language learning and processing device 10A is a device that further has a function of learning the model parameters (the output weight matrix Wout described above with reference to Fig. 4) of the storage unit 30 in the natural language processing device 10.

[0090] Fig. 10 is a block diagram showing an example of the configuration of a natural language learning and processing device 10A. The natural language learning and processing device 10A shown in Fig. 10 differs from the natural language processing device 10 shown in Fig. 3 in that the information processing device 20 is provided with a teacher data acquisition unit 23 and a learning unit 24.

[0091] The training data acquisition unit 23 acquires training data including a plurality of pairs of medical text, context information, and positive or negative labels for predetermined words contained in the medical text. The learning unit 24 learns an output weight matrix by referring to the training data acquired by the training data acquisition unit 23. This allows each component of the output weight matrix to be learned as a parameter of the prediction model.

[0092] As described above, the training data includes medical texts. These medical texts are texts recorded as text data in electronic medical records, and an example of such text data is the text data 151 shown in FIG.

[0093] As described above, the training data also includes context information. This context information is structured information included in the structured electronic medical record, and is information other than text data. An example of context information is information indicating the job title of the person who wrote the medical text.

[0094] As described above, the training data also includes positive or negative labels for specific words contained in the medical text. The specific words may be nouns in the text, and the positive or negative labels indicate whether the specific words are used in the medical text with a negative or positive meaning.

[0095] Examples of these predetermined words and positive or negative labels include "cough," "tremors," "chills," and "cloudy urine" in output column 152 shown in FIG. 8, along with the "+" or "-" labels attached to those words.

[0096] The teacher data is generated, for example, as follows: Medical text recorded as text data in a patient's electronic medical record is acquired by the teacher data acquisition unit 23, and words (e.g., nouns) to be labeled are extracted from the text data. The word extraction is performed, for example, by the teacher data acquisition unit 23 performing a morphological analysis of the text data. Context information corresponding to this text data is also acquired by the teacher data acquisition unit 23.

[0097] The acquired text data sentences and extracted words are each displayed on the display of the output unit 43. For example, an operator of the natural language learning and processing device 10A reads the medical sentences and determines whether each of the extracted words is used with a negative or positive meaning. Then, the operator of the natural language learning and processing device 10A assigns a positive label (e.g., "+") or a negative label (e.g., "-") to each word via the input unit 42.

[0098] Thereafter, the medical text recorded as text data in the electronic medical record of another patient is acquired by the training data acquisition unit 23, and the same process is performed. By repeating this process, training data including multiple pairs of medical text, context information, and positive or negative labels for predetermined words contained in the medical text is generated.

[0099] The above-described work for generating teacher data is merely an example and does not limit the present exemplary embodiment. Furthermore, the term "teacher data" in this specification does not have any other limitations than that it is data that is referenced for updating (learning) model parameters. Instead of the term "teacher data" in this specification, terms such as "learning data" or "reference data" may be used.

[0100] After the training data having a sufficient number of pairs is generated, machine learning is performed by the learning unit 24. That is, the learning unit 24 refers to the training data and learns a prediction model that represents the correlation between the medical text and context information and the positive or negative labels for predetermined words contained in the medical text.

[0101] In this case, machine learning is performed by updating the model parameters of the prediction model so that the difference between the positive or negative label for a predetermined word output by the prediction model and the positive or negative label for a predetermined word included in the training data becomes smaller. Here, the prediction model is a model using the ESN described above with reference to Figure 4, and the model parameters are the output weight matrix Wout described above.

[0102] <Learning process flow by the natural language learning and processing device 10A> The flow of the learning process executed by the natural language learning and processing apparatus 10A configured as above will be described with reference to Fig. 11. Fig. 11 is a flowchart showing the flow of the learning process.

[0103] In step S101, the training data acquisition unit 23 acquires training data including a medical text, a plurality of pairs of context information, and positive or negative labels for predetermined words included in the medical text.

[0104] In step S102, the learning unit 24 executes a parameter update process for learning an output weight matrix by referring to the training data acquired in step S101. Note that the parameter update process in step S102 is executed multiple times according to the number of pairs included in the training data acquired in step S101.

[0105] As a result, each component of the output weight matrix is ​​learned as a parameter of the prediction model.

[0106] FIG. 12 is a flowchart illustrating a detailed example of the parameter update process in step S102 of FIG.

[0107] In step S121, the learning unit 24 generates a token string from the text data included in the acquired training data.

[0108] In step S122, the learning unit 24 multiplies each token in the token sequence by an input weight matrix, where the input u(n) is multiplied by the input weight matrix Win, as described above with reference to FIG.

[0109] In step S123, the learning unit 24 multiplies the context information vector by the context weight matrix, that is, the context vector c is multiplied by the context weight matrix Wc, as described above with reference to FIG.

[0110] In step S124, the learning unit 24 multiplies the input vector, which includes the product of the input weight matrix and the token and the product of the context weight matrix and the context information vector, by the connection weight matrix. At this time, as described above with reference to Fig. 4, the input vector is multiplied by the connection weight matrix W. Note that each input vector corresponding to each token in the token sequence is multiplied by the connection weight matrix W.

[0111] In step S125, the learning unit 24 executes an output weight matrix learning process, whereby each element of the output weight matrix, which serves as a model parameter, is calculated and updated with reference to the teacher data.

[0112] In this way, the parameter update process is performed.

[0113] FIG. 13 is a flowchart illustrating a detailed example of the output weight matrix learning process in step S125 of FIG.

[0114] In step S141, the learning unit 24 acquires a high-dimensional feature vector of a predetermined word. Here, the high-dimensional feature vector is each high-dimensional feature vector generated by multiplying each input vector by a connection weight matrix in the process of step S124. At this time, a high-dimensional feature vector corresponding to a token generated from a predetermined word labeled with a label included in the training data is acquired.

[0115] In step S142, the learning unit 24 generates an output vector. At this time, the learning unit 24 acquires positive or negative labels attached to predetermined words included in the training data, and generates an output vector by vectorizing those words and labels.

[0116] In step S143, the learning unit 24 updates each element of the output weight matrix so as to reduce the difference between the high-dimensional feature vector acquired in the process of step S141 and the output vector generated in the process of step S142.

[0117] In step S144, the learning unit 142 updates the model parameters in the storage unit 30 using the components of the output weight matrix calculated in the process of step S143.

[0118] The processes of steps S141 to S144 are executed in accordance with the number of predetermined words.

[0119] 11, the parameter update process in step S102 is executed multiple times depending on the number of pairs included in the training data acquired in step S101. Therefore, each component of the output weight matrix is ​​also repeatedly updated multiple times. In this way, each component of the output weight matrix is ​​updated, and the learning of the output weight matrix progresses.

[0120] In this way, the output weight matrix learning process is executed, and the learning process shown in FIG. 11 is completed.

[0121] When the learning process is completed, the components stored as model parameters in storage unit 30 can be used as the output weight matrix Wout described above with reference to Fig. 4. That is, the stored model parameters are read out by output weight matrix multiplication unit 83 of output sequence generation unit 22 and multiplied by the high-dimensional feature vector in the output sequence generation process of Fig. 6.

[0122] In this way, it is possible to learn whether a given word contained in a medical text is used with a negative or positive meaning, and therefore, for example, it becomes possible to predict an output sequence 152 from text data 151 shown in FIG. 8.

[0123] In the above-described learning process, learning may be performed by adjusting hyperparameters as appropriate. For example, hyperparameters such as the spectrum radius, leak rate, input scaling, reservoir size, and transient response period may be adjusted by the learning unit 24.

[0124] <Natural language learning and processing device and the effects of the learning process> In this manner, the natural language learning and processing device and learning process according to this exemplary embodiment can learn correlations between medical text and context information and the (positive or negative) meaning of certain words in the medical text.

[0125] In other words, since the training data includes not only words but also medical text and context information, machine learning can be performed taking into account the order in which specific words appear in text and their relationship with the context information.

[0126] In addition, in this machine learning, model parameters of the prediction model using the ESN described above with reference to FIG. 4 are learned.

[0127] As mentioned above, in an ESN, only the output weight matrix is ​​trained, and the input weight matrix and connection weight matrix are pre-defined and do not need to be trained. This significantly reduces the computational cost during training compared to a typical RNN, and makes it possible to make predictions using a predictive model without using an extremely high-performance information processing device.

[0128] In addition, by using the learned model parameters, it is possible to predict the meaning (positive or negative) of each word in a medical sentence, so that the meaning of each word can be reflected in the sentence vector generated by adding, averaging, combining, etc. each of the output vectors in the output column.

[0129] For example, traditional language processing models such as Word2vec do not take into account the order in which words appear in a sentence, so sentence vectors for sentences containing the same word often end up being similar. However, the same word can be used in either a positive or negative sense depending on the sentence.

[0130] Figure 14 shows examples of medical texts in which the same words are used. Here, medical text 111 and medical text 112 contain the same words (nouns) such as "patient A," "drug A," and "administration," and the adjectives "restrictive" and "restrictive" are also similar. However, in terms of administering drug A to patient A, medical text 111 and medical text 112 have almost opposite meanings.

[0131] According to exemplary embodiments 1 to 3 described in this specification, for example, an output sequence can be generated so that the similarity between the sentence vector of medical sentence 111 and the sentence vector of medical sentence 112 is small.

[0132] Exemplary Embodiment 4 Next, a third exemplary embodiment of the present invention will be described in detail with reference to the drawings. Note that components having the same functions as those described in the first to third exemplary embodiments are denoted by the same reference numerals, and their description will be omitted as appropriate.

[0133] <Configuration of natural language learning device 10B> The configuration of a natural language learning device 10B according to this exemplary embodiment will be described with reference to FIG.

[0134] Natural language learning device 10B differs from natural language learning and processing device 10A described above with reference to Fig. 10 in that it does not have a function related to generating an output string, but is configured to have a function related to learning model parameters. That is, natural language learning device 10B does not have acquisition unit 21 and output string generation unit 22 provided in information processing device 20. The other configuration of natural language learning device 10B is the same as that of natural language learning and processing device 10A, and detailed description thereof will be omitted.

[0135] In natural language learning device 10B, the learning process described with reference to FIGS. 11 to 13 is executed, and the model parameters in storage unit 30 are learned. After learning is executed with reference to a sufficient number of sets of teacher data, the model parameters stored in storage unit 30 of natural language learning device 10B are stored in a storage medium such as a USB memory. The model parameters stored in the storage medium are then used by another device (e.g., natural language processing device 10 of FIG. 3). Alternatively, the model parameters of natural language learning device 10B may be transferred to another device via a network.

[0136] <Effects of natural language learning devices> In this manner, the natural language learning device according to this exemplary embodiment can learn model parameters for predicting the meaning (positive or negative) of a given word in a medical text from the medical text and context information.

[0137] In this way, for example, it is possible to provide model parameters obtained by learning with reference to different training data. Alternatively, it is also possible to provide model parameters learned by adjusting hyperparameters differently. Examples of hyperparameters include the spectral radius of the ESN, leak rate, input scaling, reservoir size, and transient response period. Using different model parameters makes it possible to perform optimal predictions depending on, for example, the characteristics of the patient or the type of disease.

[0138] Exemplary Embodiment 5 In this exemplary embodiment, examples of information that may be used as context information will be further described.

[0139] For example, the context information may be information representing the job title of the person who wrote the medical document. For example, if the job title of the person who wrote the electronic medical record is recorded in the electronic medical record, the recorded job title can be used as information representing the job title of the person who wrote the medical document.

[0140] Alternatively, a code representing a job title such as "doctor," "nurse," "paramedic," "physiotherapist," "nutritionist," etc. may be used as information representing the job title of the person who wrote the medical text.

[0141] For example, paramedics, physical therapists, and nutritionists often use different medical equipment than doctors and nurses, and often describe patients' conditions with different focus than doctors and nurses. There are also set phrases used in medical writing depending on the job title of each professional, such as doctors, nurses, paramedics, physical therapists, and nutritionists.

[0142] In this way, even if the words used in a medical text are the same, the text may have different meanings depending on the position of the person who wrote the medical text. Also, if medical texts are trained according to the position of the person who wrote them, accurate predictions can be made with a relatively small amount of training.

[0143] As another example, the context information may be the name of a field in an electronic medical record where a medical document is recorded. For example, if the electronic medical record has fields in which text should be entered, such as "Patient's Complaint," "Examination Results," "Follow-up," etc., the names of these fields may be used as context information. Alternatively, a code representing the name of a field may be used as the name of the field where the medical document is recorded.

[0144] For example, the contents described in the "Patient's Complaint" are the patient's subjective symptoms, while the contents described in the "Examination Results" are objective symptoms based on test results. Also, even if the patient's understanding of "severe illness" is entered as the chief complaint, the "severe illness" used in the examination results means a condition requiring a cardiopulmonary bypass or artificial respirator, which is often very different from the patient's understanding.

[0145] In this way, even if the words used in a medical sentence are the same, the sentence may have a different meaning depending on the field the medical sentence is recorded in. Also, by learning medical sentences according to the recorded field, accurate predictions can be made with a relatively small amount of learning.

[0146] Here, an example of context information has been described as the name of the field in which the medical document is recorded, but it may also be the name of the tab, record, table, etc. in which the medical document is recorded depending on the structure of the electronic medical record.

[0147] In the exemplary embodiments presented herein, context information is used to predict output sequences and / or train output vectors, allowing accurate predictions to be made with a relatively small amount of training.

[0148] [Software implementation example] Some or all of the functions of the information processing device 20, the natural language processing device 10, the natural language learning and processing device 10A, and the natural language learning device 10B may be realized by hardware such as an integrated circuit (IC chip), or by software.

[0149] In the latter case, the information processing device 20, the natural language processing device 10, the natural language learning and processing device 10A, and the natural language learning device 10B are realized by, for example, a computer that executes instructions of a program, which is software that realizes each function. An example of such a computer (hereinafter referred to as computer C) is shown in FIG. 16.

[0150] The computer C includes at least one processor C1 and at least one memory C2. The memory C2 stores a program P for causing the computer C to operate as an information processing device 20, a natural language processing device 10, a natural language learning and processing device 10A, or a natural language learning device 10B. In the computer C, the processor C1 reads and executes the program P from the memory C2, thereby realizing each function of the information processing device 20, the natural language processing device 10, the natural language learning and processing device 10A, or the natural language learning device 10B.

[0151] The processor C1 may be, for example, a central processing unit (CPU), a graphics processing unit (GPU), a digital signal processor (DSP), a micro processing unit (MPU), a floating point number processing unit (FPU), a physics processing unit (PPU), a microcontroller, or a combination thereof. The memory C2 may be, for example, a flash memory, a hard disk drive (HDD), a solid state drive (SSD), or a combination thereof.

[0152] The computer C may further include a RAM (Random Access Memory) for expanding the program P during execution and for temporarily storing various data. The computer C may also include a communication interface for transmitting and receiving data to and from other devices. The computer C may also include an input / output interface for connecting input / output devices such as a keyboard, mouse, display, and printer.

[0153] Furthermore, the program P can be recorded on a non-transitory tangible recording medium M that can be read by the computer C. Such a recording medium M can be, for example, a tape, a disk, a card, a semiconductor memory, or a programmable logic circuit. The computer C can acquire the program P via such a recording medium M. The program P can also be transmitted via a transmission medium. Such a transmission medium can be, for example, a communication network or broadcast waves. The computer C can also acquire the program P via such a transmission medium.

[0154] [Appendix 1] The present invention is not limited to the above-described embodiments, and various modifications are possible within the scope of the claims. For example, embodiments obtained by appropriately combining the technical means disclosed in the above-described embodiments are also included in the technical scope of the present invention.

[0155] [Appendix 2] Some or all of the above-described embodiments can also be described as follows: However, the present invention is not limited to the following described aspects.

[0156] (Appendix 1) an acquisition means for acquiring a token sequence obtained from a medical document in an electronic medical record and a context information vector obtained from context information in the electronic medical record; an output sequence generation means for executing an output sequence generation process for generating an output sequence from the token sequence and the context information vector, the output sequence generation process including a process for converting the token sequence into a high-dimensional feature vector having a dimension higher than the sum of the dimensions of the token sequence and the context information vector; 1. An information processing device comprising:

[0157] (Appendix 2) The acquisition means a token sequence generation means for generating the token sequence by converting each word included in the medical text into a token by embedding it in a predefined first vector space; and a context information vector generating means for extracting predetermined context information from the electronic medical record and embedding the context information in a predefined second vector space to generate the context information vector. 2. The information processing device according to claim 1,

[0158] (Appendix 3) The output sequence generation means An input vector including a product of a predetermined input weight matrix and the token and a product of a predetermined context weight matrix and the context information vector is multiplied by a predetermined connection weight matrix to convert the input vector into the high-dimensional feature vector. 3. The information processing device according to claim 2,

[0159] (Appendix 4) The output sequence generating means further The output sequence is generated by multiplying the high-dimensional feature vector by an output weight matrix obtained by prior learning. 4. The information processing device according to claim 3,

[0160] (Appendix 5) The system further includes a learning unit that learns the output weight matrix by referring to training data including a plurality of pairs of medical text, context information, and positive or negative labels for predetermined words included in the medical text. 5. The information processing device according to claim 4,

[0161] (Appendix 6) The context information is information indicating the position of the person who wrote the medical text. 6. The information processing device according to any one of Supplementary Notes 1 to 5.

[0162] (Appendix 7) The context information is the name of the field in the electronic medical record where the medical document is recorded. 6. The information processing device according to any one of Supplementary Notes 1 to 5.

[0163] (Appendix 8) obtaining a token sequence obtained from medical text in an electronic medical record and a context information vector obtained from context information in the electronic medical record; and performing an output sequence generation process for generating an output sequence from the token sequence and the context information vector, the output sequence generation process including a process for converting the token sequence into a high-dimensional feature vector having a dimension higher than the sum of the dimensions of the token sequence and the dimensions of the context information vector. 1. An information processing method comprising:

[0164] (Appendix 9) The output sequence generation process includes: generating the token sequence by converting each word included in the medical text into a token by embedding it in a predefined first vector space; extracting predetermined context information from the electronic medical record and generating the context information vector by embedding the context information in a predefined second vector space. 9. The information processing method according to claim 8,

[0165] (Appendix 10) The output column processing further comprises: An input vector including a product of a predetermined input weight matrix and the token and a product of a predetermined context weight matrix and the context information vector is multiplied by a predetermined connection weight matrix to convert the input vector into the high-dimensional feature vector. 10. The information processing method according to claim 9,

[0166] (Appendix 11) The output column processing further comprises: The output sequence is generated by multiplying the high-dimensional feature vector by an output weight matrix obtained by prior learning. 9. The information processing method according to claim 8,

[0167] (Appendix 12) The method further includes training the output weight matrix by referring to training data including a plurality of pairs of predetermined words included in the medical text and context information and positive or negative labels related to the predetermined words included in the medical text. 12. The information processing method according to claim 11.

[0168] (Appendix 13) The context information is information indicating the position of the person who wrote the medical text. 13. The information processing method according to any one of appendices 8 to 12.

[0169] (Appendix 14) The context information is the name of the field in the electronic medical record where the medical document is recorded. 13. The information processing method according to any one of appendices 8 to 12.

[0170] (Appendix 15) Computer, an acquisition means for acquiring a token sequence obtained from a medical document in an electronic medical record and a context information vector obtained from context information in the electronic medical record; and an information processing device that functions as an information processing apparatus comprising: an output sequence generation means that executes an output sequence generation process that generates an output sequence from the token sequence and the context information vector, the output sequence generation process including a process of converting the token sequence into a high-dimensional feature vector having a dimension higher than the sum of the dimension of the token sequence and the dimension of the context information vector. A program characterized by:

[0171] (Appendix 16) The acquisition means a token sequence generation means for generating the token sequence by converting each word included in the medical text into a token by embedding it in a predefined first vector space; and a context information vector generating means for extracting predetermined context information from the electronic medical record and embedding the context information in a predefined second vector space to generate the context information vector. 16. The program according to claim 15,

[0172] (Appendix 17) The output sequence generation means An input vector including a product of a predetermined input weight matrix and the token and a product of a predetermined context weight matrix and the context information vector is multiplied by a predetermined connection weight matrix to convert the input vector into the high-dimensional feature vector. 17. The program according to claim 16,

[0173] (Appendix 18) The output sequence generating means further The output sequence is generated by multiplying the high-dimensional feature vector by an output weight matrix obtained by prior learning. 18. The information processing device according to claim 17,

[0174] (Appendix 19) The system further includes a learning unit that learns the output weight matrix by referring to training data including a plurality of pairs of medical text, context information, and positive or negative labels for predetermined words included in the medical text. 19. The information processing device according to claim 18,

[0175] (Appendix 20) The context information is information indicating the position of the person who wrote the medical text. 20. The information processing device according to any one of Supplementary Notes 15 to 19.

[0176] (Appendix 21) The context information is the name of the field in the electronic medical record where the medical document is recorded. 20. The information processing device according to any one of Supplementary Notes 15 to 19.

[0177] [Appendix 3] A part or all of Exemplary Embodiment 4 can also be described as follows, but the present invention is not limited to the following described aspects.

[0178] (Appendix 22) By referring to training data including a plurality of pairs of medical text and context information of the electronic medical record and positive or negative labels for predetermined words included in the medical text, The system includes a learning means for learning an output weight matrix that represents the correlation between a high-dimensional feature vector obtained by converting a token sequence obtained from a medical text and a context information vector obtained from context information, the high-dimensional feature vector having a dimension higher than the sum of the dimensions of the token sequence and the context information vector, and an output sequence made up of output vectors that vectorize predetermined words included in the medical text and positive or negative labels related to the words. A learning device characterized by:

[0179] (Appendix 23) By referring to training data including a plurality of pairs of medical text and context information of the electronic medical record and positive or negative labels for predetermined words included in the medical text, The method includes learning an output weight matrix that represents the correlation between a high-dimensional feature vector obtained by converting a token sequence obtained from a medical text and a context information vector obtained from context information, the high-dimensional feature vector having a dimension higher than the sum of the dimensions of the token sequence and the context information vector, and an output sequence consisting of output vectors that vectorize predetermined words included in the medical text and positive or negative labels related to the words. A learning method characterized by:

[0180] (Appendix 24) Computer, By referring to training data including a plurality of pairs of medical text and context information of the electronic medical record and positive or negative labels for predetermined words included in the medical text, The device functions as a learning device having a learning means for learning an output weight matrix that represents the correlation between a high-dimensional feature vector obtained by converting a token sequence obtained from a medical text and a context information vector obtained from context information, the high-dimensional feature vector having a dimension higher than the sum of the dimensions of the token sequence and the context information vector, and an output sequence made up of output vectors that vectorize predetermined words included in the medical text and positive or negative labels related to the words. A program characterized by:

[0181] [Appendix 3] Some or all of the above-described embodiments can also be expressed as follows.

[0182] at least one processor, the processor comprising: an acquisition process for acquiring a token sequence obtained from the medical text of an electronic medical record and a context information vector obtained from the context information of the electronic medical record; An output sequence generation process is executed to generate an output sequence from the token sequence and the context information vector, the output sequence generation process including a process of converting the token sequence into a high-dimensional feature vector having a dimension higher than the sum of the dimensions of the token sequence and the dimensions of the context information vector.

[0183] The information processing device may further include a memory that stores a program for causing the processor to execute the acquisition process and the output sequence generation process. The program may also be recorded on a computer-readable, non-transitory, tangible recording medium. [Explanation of symbols]

[0184] 10 Natural Language Processing 10A Natural Language Learning and Processing Device 10B Natural Language Learning Device 20 Information processing equipment 21 Acquisition Department 22 Output column generation unit 23 Teacher data acquisition unit 24 Learning Department 30 Storage section 41 Communications Department 42 Input section 43 Output section 61 Token sequence generation unit 62 Context information vector generation unit 81 Input weight matrix multiplication unit 82 Connection weight matrix multiplication unit 83 Output weight matrix multiplication unit

Claims

1. Token sequences obtained from medical text in electronic medical records and context information of the electronic medical records and an acquisition means for acquiring a context information vector obtained from an output sequence generation process for generating an output sequence from the token sequence and the context information vector; The dimension of the token sequence is larger than the sum of the dimension of the context information vector. The output column generation process includes converting the output column into a high-dimensional feature vector. and a force train generating means.

1. An information processing device comprising:

2. The acquisition means Embedding each word contained in the medical text into a predefined first vector space. a token sequence generating means for generating the token sequence by converting the token sequence into tokens by Extracting predetermined context information from the electronic medical record, The context information is then embedded in a second predefined vector space. a context information vector generating means for generating a vector 2. The information processing apparatus according to claim 1, wherein:

3. The output sequence generation means A predetermined input weight matrix is ​​multiplied by the token and a predetermined context is calculated. The input vector includes a product of a weight matrix and the context information vector, and a predetermined The input vector is converted into the high-dimensional feature vector by multiplying it by the connection weight matrix. R 3. The information processing apparatus according to claim 2, wherein:

4. The output sequence generating means further The high-dimensional feature vector is multiplied by an output weight matrix obtained by prior learning, Generate the output sequence 4. The information processing apparatus according to claim 3,

5. Medical text and context information and positive responses regarding specific words contained in the medical text Or, the output weight matrix is ​​learned by referring to training data including a plurality of pairs with negative labels. It also has a learning section 5. The information processing apparatus according to claim 4,

6. The context information is information indicating the position of the person who wrote the medical text.

6. The information processing apparatus according to claim 1, wherein:

7. The context information is a field in the electronic medical record where the medical document is recorded. is the name of 6. The information processing apparatus according to claim 1, wherein:

8. An information processing device, Token sequences obtained from medical text in electronic medical records and context information of the electronic medical records and obtaining a context information vector obtained from an output sequence generation process for generating an output sequence from the token sequence and the context information vector; The dimension of the token sequence is larger than the sum of the dimension of the context information vector. Execute output column generation processing, including processing to convert into a high-dimensional feature vector having high dimensions. and includes 1. An information processing method comprising:

9. Computer, Token sequences obtained from medical text in electronic medical records and context information of the electronic medical records and an acquisition means for acquiring a context information vector obtained from an output sequence generation process for generating an output sequence from the token sequence and the context information vector; The dimension of the token sequence is larger than the sum of the dimension of the context information vector. The output column generation process includes converting the output column into a high-dimensional feature vector. and a force train generating means. A program characterized by:

Images