JP2020106880A - Information processing apparatus, model generation method and program - Google Patents

Abstract

To prevent an overlearning of a frequent expression in learning text data in generating a learning model.SOLUTION: An information processing apparatus 50 that generates a model for extracting a predetermined expression from text data includes a feature generation unit 22 that divides learning text data into words and sets feature data for each word, a labeling unit 23 that gives a label for each word, a replacement unit 24 that replaces an expression having a predetermined attribute with respect to a character characteristic included in the learning text data or the feature data with a character string randomly generated, and a learning unit 25 that generates the model for extracting the predetermined expression by using the replaced data and the labeled feature data as learning data.SELECTED DRAWING: Figure 2

Description

The present disclosure relates to a method of creating learning data for machine learning.

Conventionally, for the purpose of constructing a model for automatically and highly semantically tagging words, the features of the context system for words in a learning corpus in which each word is given a semantic tag and a grammar tag. The process of extracting the function and the dependency feature function, the process of extracting the ontology feature function from the ontology, and for each of these feature functions, based on the statistics of the value of the feature function calculated for each word in the learning corpus Based on the feature selection process that calculates the mutual information of the feature function and selects the feature function with a large amount of information, and the statistics of the value of the feature function obtained for each word in the learning corpus regarding the selected feature function. , A process of calculating a statistical model that gives the likelihood that each meaning tag is given to the word based on the value of the feature function obtained for each word in the input sentence, and a model building device is proposed. (See Patent Document 1).

In addition, for the purpose of improving the accuracy of the semantic classification assignment, the feature weight table is used as a feature weight that indicates the relationship between the text feature and the semantic classification in advance, and the subsequent functional expression of the phrase and the subsequent functional expression A weight that is characterized by the difference in the meaning classification from the relevant phrase is held, and the score calculation unit searches the table using the feature of the word of interest in the text received by the input unit as a key, and the meaning classification given to the word The tag candidate and the score are calculated, and the semantic classification giving means selects the series having the maximum score sum from the series of the semantic classification tag candidates, and gives the semantic classification tag by adding the meaning classification tag to each word in the input text. It has been proposed to output the result (see Patent Document 2).

Further, for the purpose of improving the extraction accuracy of a word that is included in the dictionary but is ambiguous and a word that is not included in the dictionary, the teacher data creation device has a storage unit that stores dictionary data created manually and a tag. An input device for inputting unlearned learning target text, and a processing unit for creating tagged text from untagged learning target text based on word meaning information indicating meaning classification of words included in dictionary data. Is proposed (see Patent Document 3).

Furthermore, a unified framework using a bidirectional long short-term memory network for named entity recognition tasks in biomedical and clinical areas has been proposed (see Non-Patent Document 1).

JP, 2008-021139, A JP, 2011-129006, A International Publication 2015/177861

Sunil Kumar Sahu, Ashish Andand, "Unified Neural Architectural For Drug, Disease and Clinical Entity Recognition", https://arxiv.org/pdf/1708.447.

Conventionally, when extracting an expression having a predetermined attribute regarding a character feature from the extraction sentence data, a series labeling problem, which is a problem of assigning an appropriate label to each element (each word) of a series (word string), is solved. It is common to get burned. Then, an extraction model for that purpose is created using the learning text data.

Here, character features, character type features, and part-of-speech features are used as general features (features) when performing proper expression extraction, but there are frequent words that belong to the category to be extracted in the learning text data. In this case, the feature of the character feature strongly influences and over-learns the frequent word itself, so that an extraction model in which unknown words not included in the learning sentence data are likely to be omitted is created.

In view of the above-mentioned problems, the present disclosure aims to prevent over-learning of expressions that frequently appear in learning text data when creating a learning model.

One example of the present disclosure is an information processing apparatus for creating a model for extracting a predetermined expression from text data, in which learning text data is divided into words, and each word includes a plurality of character features including character features. Feature generation means for setting feature data about features, label assigning means for labeling each word in the feature data, and predetermined attributes for character features included in the learning text data or the feature data Replacement means for generating replaced data by replacing one or more expressions with a character string randomly generated for each expression, and the replaced data and the labeled feature data are learning data. And a learning unit that creates a model for extracting an expression having a predetermined attribute regarding a character feature from newly input extraction sentence data.

The present disclosure can be understood as an information processing device, a system, a method executed by a computer, or a program executed by a computer. Further, the present disclosure can be grasped as such a program recorded on a recording medium readable by a computer or other device, machine, or the like. Here, a computer-readable recording medium is a recording medium that stores information such as data and programs by electrical, magnetic, optical, mechanical, or chemical action and can be read by a computer. Say.

According to the present disclosure, it is possible to prevent over-learning of expressions that frequently appear in learning text data when creating a learning model.

1 is a schematic diagram showing a hardware configuration of a system according to an embodiment. It is a figure which shows the outline of a function structure of the information processing apparatus which concerns on embodiment. It is a figure which shows the whole image of the process performed by the information processing apparatus which concerns on embodiment. It is a flow chart which shows an outline of a flow of learning data creation processing concerning an embodiment. It is a flow chart which shows an outline of a flow of data preparation processing concerning an embodiment. It is a flow chart which shows the outline of the flow of the feature generation processing concerning an embodiment. It is a figure which shows the outline|summary of the characteristic data used in embodiment. It is a figure which shows the outline|summary of the feature data after a replacement used in embodiment. It is a flow chart which shows the outline of the flow of the extraction processing concerning an embodiment. It is a figure which shows the outline of a function structure of the information processing apparatus which concerns on a variation.

Hereinafter, embodiments of an information processing device, a model creating method, and a program according to the present disclosure will be described with reference to the drawings. However, the embodiments described below are examples of the embodiments, and the information processing apparatus, the model creating method, and the program according to the present disclosure are not limited to the specific configurations described below. In implementation, a specific configuration according to the embodiment may be appropriately adopted, and various improvements and modifications may be performed.

In the present embodiment, the information processing apparatus, the model creating method, and the program according to the present disclosure collect information regarding information processing security from various information sources (natural language) and extract a malware name from the collected information. An embodiment when implemented in the system will be described. It should be noted that the information processing device, the model creating method, and the program according to the present disclosure can be widely used for a technique for extracting an expression having a predetermined attribute from text data. It is not limited to the example shown in. For example, the technology according to the present disclosure is useful when creating a model for extracting a noun in a category in which a new word is likely to occur from a sentence.

<System configuration>
FIG. 1 is a schematic diagram showing a hardware configuration of a system 1 according to this embodiment. The system 1 according to the present embodiment includes an information processing device 50 communicably connected to a server (not shown) on the Internet.

The information processing device 50 includes a CPU (Central Processing Unit) 11, a ROM (Read Only Memory) 12, a RAM (Random Access Memory) 13, and an EEPROM (Electrically Erasable and Erasable and Programmable Drained Read on). The computer includes a device 14, a communication unit such as a NIC (Network Interface Card) 15, and the like. However, the specific hardware configuration of the information processing device 50 can be appropriately omitted, replaced, or added depending on the embodiment. Further, the information processing device 50 is not limited to a device having a single housing. The information processing device 50 may be realized by a plurality of devices using so-called cloud or distributed computing technology.

FIG. 2 is a diagram showing an outline of a functional configuration of the information processing device 50 according to the present embodiment. In the information processing device 50, the program recorded in the storage device 14 is read into the RAM 13 and executed by the CPU 11 to control each hardware included in the information processing device 50, thereby obtaining the sentence data. It functions as an information processing device including the unit 21, the feature generation unit 22, the label assignment unit 23, the replacement unit 24, the learning unit 25, and the extraction unit 26. In the present embodiment and other embodiments described later, each function of the information processing device 50 is executed by the CPU 11 that is a general-purpose processor, but some or all of these functions are dedicated to one or more dedicated ones. It may be executed by a processor.

The text data acquisition unit 21 collects pages including security-related articles from the Internet, extracts text from the pages, and acquires learning text data or extraction text data.

The feature generation unit 22 divides the learning sentence data into words, and sets, for each word, feature data for a plurality of features including a character feature (in the present embodiment, a character feature, a character type feature, and a part-of-speech feature).

The label assigning unit 23 attaches a label (in this embodiment, an IOB2 (Inside, outside, beginning 2) system label) to each word in the feature data.

The replacement unit 24 replaces one or a plurality of expressions having a predetermined attribute regarding a character feature included in the learning text data or the feature data with a randomly generated character string for each expression, thereby performing replacement. Generate data. When replacing the expression with a randomly generated character string, the replacement unit 24 generates a character string that matches the expression with respect to features other than the character features (part of speech feature and character type feature in the present embodiment). And replace.

The learning unit 25 uses, as learning data, the replaced data and the labeled feature data, an expression (a noun of a category such as a malware name in which a new word is likely to occur) having a predetermined attribute with respect to a character feature from the extraction text data. Create a model for extraction.

The extraction unit 26 uses the model created by the learning unit 25 to extract an expression having a predetermined attribute regarding a character feature from the newly input text data for extraction.

<Process flow>
Next, a flow of processing executed by the information processing device 1 according to the present embodiment will be described. Note that the specific contents and processing order of the processing described below are examples for implementing the present disclosure. Specific processing contents and processing order may be appropriately selected according to the embodiment of the present disclosure.

FIG. 3 is a diagram showing an overall image of processing executed by the information processing apparatus according to the present embodiment. The information processing device 50 acquires a security-related article for learning from the outside, creates learning data, and creates an extraction model by learning using the learning data. Then, the information processing device 50 uses the created extraction model to extract the malware name from the security-related article for extraction acquired from the outside. Hereinafter, the details of each process will be described using a flowchart.

FIG. 4 is a flowchart showing an outline of the flow of the learning data creation processing according to this embodiment. The process shown in this flowchart is executed at any time when it is desired to create learning data for creating a model for extracting a malware name.

In step S101, a data preparation process is executed. Details of the data preparation process will be described with reference to FIG.

FIG. 5 is a flowchart showing an outline of the flow of data preparation processing according to the present embodiment. The process shown in this flowchart shows the details of the data preparation process shown in step S101 of the learning data creation process and step S401 of the extraction process described later. The process is executed when the data preparation process is called in step S401.

In steps S201 and S202, security related articles are collected and texts are extracted. The text data acquisition unit 21 collects HTML pages including security-related articles from the Internet (step S201), and extracts texts (here, text parts in natural language excluding tags, scripts, etc.) from the HTML pages. It is extracted (step S202). Then, a process progresses to step S203.

In step S203, the data is divided for each sentence. The text data acquisition unit 21 acquires text data (learning text data or extraction text data) by dividing the text extracted in step S202 into sentence units according to a predetermined text division rule. Since the division rule into sentences varies depending on the language of the text, it is preferable to adopt an appropriate division rule according to the target language. For example, a method of dividing based on a punctuation mark, a period, a line break, etc. used in the language can be adopted, but the division rule that can be adopted here is not limited. After that, the processing shown in this flowchart ends.

The description is returned to the learning data creation process of FIG. When the data preparation process of step S101 is completed, the process proceeds to step S102.

In step S102, the malware name is extracted from the learning text data and a malware name dictionary is created. Here, the process of extracting the malware name from the learning text data may be performed by analysis using an analysis program, or may be performed by a person who has read the text. The malware name extracted from the learning text data is registered in the malware name dictionary. As the malware name dictionary, it is possible to use existing dictionary data other than the one created in this step. If only existing dictionary data is used, this step may be omitted. After that, the process proceeds to step S103.

In step S103, a feature generation process is executed. Details of the feature generation processing will be described with reference to FIG.

FIG. 6 is a flowchart showing an outline of the flow of the feature generation processing according to this embodiment. The process shown in this flowchart shows the details of the feature generation process shown in step S103 of the learning data creation process and step S402 of the extraction process described later. The process is executed when the feature generation process is called in step S402.

In step S301, the sentence data is divided into words. The feature generation unit 22 divides the sentence data (learning sentence data or extraction sentence data) into words according to a predetermined word division rule. Since the division rule into words varies depending on the language of the sentence, it is preferable to adopt an appropriate division rule according to the target language. For example, a method of dividing with a space character as a reference, a method of dividing with a character type as a reference, a method of identifying and dividing a word according to a dictionary, and the like can be adopted, but the dividing rule that can be adopted here is not limited. Then, a process progresses to step S302.

In step S302 and step S303, a part-of-speech tag and a character type tag are added. The feature generation unit 22 sets feature data for a plurality of features including a character feature (in the present embodiment, a character feature, a character type feature, and a part-of-speech feature) for each word obtained in step S301.

FIG. 7 is a diagram showing an outline of the feature data used in this embodiment. The characteristic data is created for each target word in the sentence. The feature data shown in FIG. 7 is learning data using character (word) features, part-of-speech (pos) features, and character type (shape) features of nine words from four to four after the target word as features. It represents.

In the character feature column in the feature data, the words in the sentence are shown as they are. In the example shown here, 9 words “the global spread of WannaCry confirmed on or after” centered on the target word “WannaCry” are extracted from the sentence data and stored in the learning data for each word. .. Then, each word is assigned a code from -4 to +4, which indicates the relative position (distance) from the target word (see the leftmost column in the drawing). More specifically, the target word is indicated by “0”, the word after the target word is indicated by the distance (the number of words) from the target word, and the word before the target word is a minus sign. Is indicated by the distance from the target word (the number of words).

The part-of-speech feature is a feature indicating the part-of-speech of a word. The feature generation unit 22 estimates the part-of-speech of each word by searching a dictionary in which the word and the combination of the part-of-speech of the word are accumulated, analyzes using a grammar analysis program, and the like, and provides information indicating the part-of-speech for each word. Attached. In the example shown in FIG. 7, DT is a determiner, JJ is an adjective, NN is a singular noun, IN is a preposition or subordinate conjunction, NNP is a singular proper noun, VBD is a past tense verb, CC indicates a coordination conjunction. In the present embodiment, tagging is performed using the English part-of-speech tagging model of Stanford Core NLP, but the standard used for tagging is not limited to the examples in this disclosure.

The character type feature is a feature about the shape of a word, and an uppercase letter X and a lowercase letter x are used to represent the shape of the word. In the present embodiment, when the target word has five or more characters, the shapes of the first and last two characters are retained, and the shapes of the characters between them are collectively represented as one shape. There is. The feature generation unit 22 specifies a character type feature of each word by searching a dictionary in which the word and a combination of character type features of the word are accumulated, and attaches information indicating the character type to each word. In the present embodiment, the character type information is tagged using the library of Stanford Core NLP, but the algorithm used for tagging is not limited to the example in the present disclosure.

The characteristic data is generated for each word in the text data. When the generation of the characteristic data in step S302 and step S303 is completed, the process shown in this flowchart ends.

The description is returned to the learning data creation process of FIG. When the feature generation process of step S103 is completed, the process proceeds to step S104.

In step S104, the correct answer label is given by the character string matching based on the dictionary. The label assigning unit 23 attaches a label (tag) to each word in the characteristic data. More specifically, the labeling unit 23 performs character string matching between each word extracted from the learning sentence data in step S301 and the malware name registered in the malware name dictionary created in step S102. The word of the malware name included in the text data for learning is identified by, and the correct answer label is attached to the word identified as the malware name in the feature data. In the present embodiment, an example is shown in which a label is attached according to the IOB2 method and "B-NAME" is used as the correct answer label (see FIG. 7), but other methods may be used for the label. Further, in the present embodiment, the malware name dictionary prepared in step S102 is used as the dictionary for identifying the words of the malware name from the learning text data, but the existing dictionary data is used as the malware name dictionary. You may be asked. Then, a process progresses to step S105.

In step S105, the malware name is replaced with a random character string. The replacement unit 24 replaces the expressions (words in the present embodiment), which are included in the learning text data and the feature data and are given the correct answer label in step S104, with a character string that is randomly generated for each expression. By doing so, the replaced data is generated. That is, when the same malware name appears multiple times in the learning text data or feature data to be replaced, the replacement unit 24 replaces each of the multiple malware names that appeared multiple times with different random character strings. To do. Specifically, when "WannaCry" appears multiple times, the replacement unit 24 determines that the first "WannaCry" is "GpostVrt", the second "WannaCry" is "VrettWpo", and so on. Replace with different character strings. Therefore, for example, if there is learning text data "The WannaCry ransomware attack was a May 2017 worldwide cyberattack by the WannaCry ransomware cryptoworm,...", this text is replaced by "The GpostVrt ransomware attack was a May 2017 worldwide. cyberattack by the VrettWpo ransomware cryptoworm, …”.

Further, the replacing unit 24 generates and replaces a character string that matches the expression with respect to features other than the character features (in the present embodiment, part-of-speech features and character type features). Specifically, in the present embodiment, since the target expression is a malware name (proper noun), the replacement unit 24 generates a character string as a proper noun when generating a random character string. Here, it may be confirmed that the generated random character string does not overlap with a word related to another part of speech, another common noun, or the like. Regarding the character type feature, the replacement unit 24 randomly generates and replaces a character string such that the character type feature of the replacement target character string and the character type feature of the replaced character string are not different. Specifically, when the target expression is "WannaCry", the character type feature is "XxXxxx", so that the substitution unit 24 uses the character strings such as "GpostVrt" and "VrettWpo" so that the character type features are the same. To generate.

FIG. 8 is a diagram showing an outline of the feature data after replacement, which is used in this embodiment. Referring to the feature data after replacement shown in FIG. 8, as compared with the feature data before replacement shown in FIG. 7, the word “WannaCry”, which is a word labeled with the correct answer, is a random character string “ You can see that it has been replaced with "NvxobOgr". Further, although the word is replaced with respect to the character feature, it can be seen that the part-of-speech feature (singular proper noun) and the character type feature (“XxXxxx”) are the same as before the replacement.

When the replacement is completed for all the malware names, the process shown in this flowchart ends.

When the learning text data is created by the processing described above, the learning unit 25 uses the replaced and labeled feature data (see FIG. 8) as learning data to extract the malware name from the text data. Create a model. As a specific method of creating a model, a method of creating a model for CRF (Conditional Random Fields) or the like used in sequence labeling may be used, and a detailed description thereof will be omitted.

In the creation of the learning model using the CRF, the extraction accuracy is improved by optimizing the weight λ for each element of the feature vector. For this reason, in the conventional learning method, if the same malware name appears frequently, the character features of the malware name are over-learned (the weight λ of the character feature of the known malware name is too large). In the created model, there was a risk of omission of extraction for the new malware name. On the other hand, since the model created in the present embodiment is a model created using learning data in which the malware name is replaced with a random character string, overlearning the character features of the known malware name. It becomes a model that can prevent omission of extraction of not only known malware names but also new malware names not in the dictionary without (the character feature weight λ of the known malware name becomes excessive).

FIG. 9 is a flowchart showing an outline of the flow of the extraction process according to this embodiment. The process shown in this flowchart is executed at any time when it is desired to acquire the malware name from the Internet.

In steps S401 and S402, a data preparation process and a feature generation process are executed. The details of the data preparation process have been described with reference to FIG. 5, and the details of the feature generation process have been described with reference to FIG. Then, a process progresses to step S403.

In step S403, the malware name is extracted. Using the model created by the learning unit 25, the extraction unit 26 extracts a new malware name that is not in the dictionary in addition to the known malware name from the extraction sentence data newly obtained in step S401. As for a specific method of expression extraction using a model, an extraction method using a model for CRF or the like used in sequence labeling may be used, and a description thereof will be omitted. After that, the processing shown in this flowchart ends.

<First variation>
According to the technique of the present disclosure, the number of learning text data and feature data used for learning can be increased, and the learning effect can be further improved. That is, in the present embodiment, since the data used for learning is the data in which the word with the correct answer is replaced with a random character string, the same pre-replacement data is replaced with a plurality of replacements with different character strings. Later data can be prepared. Specifically, in step S105, the replacement unit 24 creates a plurality of pieces of data in which, with respect to certain learning text data and feature data, different random character strings are replaced when the character strings are replaced.

Therefore, in the variation, the replacement unit 24 randomly generates, for one learning text data or feature data, the first replaced data in which the expression is replaced with the randomly generated first character string. And second replaced data in which the expression is replaced with the second character string. Then, the learning unit 25 creates a model for extraction using the first replaced data and the second replaced data.

For example, when replacing the learning sentence data and feature data including the above "the global spread of WannaCry confirmed on or after", by replacing "WannaCry" with "NvxobOgr", "the global spread of NvxobOgr confirmed on or after, the first post-substitution data, and by replacing the "WannaCry" with "TrssePnh", the second post-substitution data with "the global spread of TrssePnh confirmed on or after", and so on. It is possible to prepare a plurality of different learning text data and characteristic data based on one learning text data and characteristic data and use them for learning.

<Second variation>
The information processing apparatus according to the present disclosure has the effect of preventing omission of extraction of a new malware name as described above, but on the other hand, there is a possibility that a word that is not a malware name will be erroneously extracted. Therefore, as a variation of this embodiment, it is possible to additionally register words that are not malware names in a dictionary and perform filtering.

FIG. 10 is a diagram showing an outline of a functional configuration of the information processing device 50b according to this variation. In this variation, the information processing device 50b uses the recording unit 27 that records an expression that does not have a predetermined attribute (in this embodiment, a word that is not a malware name) and the expression extracted by the extraction unit 26 from the recording unit. The filtering unit 28 that excludes expressions (in this embodiment, words that are not malware names) that do not have a predetermined attribute and that are recorded by 27 are further included.

<Effect>
By using the system disclosed in the present embodiment, the user can reduce extraction omission of articles regarding new malware when identifying articles regarding predetermined malware from a large number of security-related articles existing on the Internet. it can. This is because the technology according to the present disclosure makes it possible to prevent over-learning of expressions that frequently appear in learning text data when creating a learning model. The technology of the present disclosure is not limited to the embodiment described above, and can be used when creating a model for extracting a noun in a category in which a new word is likely to occur from a sentence.

1 system 50 information processing device

Claims (12)

An information processing device for creating a model for extracting a predetermined expression from text data,
Feature generation means for dividing the learning sentence data into words, and setting feature data for a plurality of features including character features for each word,
In the characteristic data, a labeling unit that labels each word,
Replaced data is generated by replacing, for each expression, one or a plurality of expressions that have a predetermined attribute regarding the character feature included in the learning text data or the feature data, with a randomly generated character string. Replacement means to
Learning means for creating a model for extracting an expression having a predetermined attribute for a character feature from newly input extraction sentence data, using the replaced data and the labeled feature data as learning data,
An information processing apparatus including. The replacement means, when replacing the expression with a randomly generated character string, generates and replaces a character string that matches the expression for features other than character features.
The information processing apparatus according to claim 1. The feature generation means sets feature data for a plurality of features further including a character feature, a character type feature, and a part-of-speech feature,
The information processing apparatus according to claim 1. The replacement means, when replacing the expression with a randomly generated character string, generates and replaces a character string that matches the target expression with respect to the part-of-speech feature and the character type feature.
The information processing device according to claim 3. The replacement means, with respect to one of the learning sentence data or the feature data, first replaced data in which the expression is replaced with a first character string that is randomly generated, and second randomly generated data. A second replaced data in which the expression is replaced with a character string, and
The learning means creates the model using the first replaced data and the second replaced data.
The information processing apparatus according to any one of claims 1 to 4. The expression having the predetermined attribute is a noun of a category in which a new word is likely to occur,
The information processing apparatus according to any one of claims 1 to 5. The expression having the predetermined attribute is a malware name,
The information processing device according to claim 6. Further comprising text data acquisition means for collecting pages including security related articles from the Internet, extracting text from the pages, and acquiring the learning text data or the extraction text data.
The information processing device according to claim 7. Further comprising extraction means for extracting an expression having the predetermined attribute from the extraction sentence data using the model created by the learning means,
The information processing apparatus according to any one of claims 1 to 8. Recording means for recording an expression not having the predetermined attribute,
Filtering means for excluding, from the expressions extracted by the extracting means, expressions that do not have the predetermined attribute, which are recorded by the recording means.
The information processing device according to claim 9. A computer for creating a model for extracting a predetermined expression from text data,
A feature generation step of dividing the learning sentence data into words, and setting feature data for a plurality of features including character features for each word,
In the characteristic data, a labeling step of labeling each word,
Replaced data is generated by replacing, for each expression, one or a plurality of expressions that have a predetermined attribute regarding the character feature included in the learning text data or the feature data, with a randomly generated character string. Replacement step,
A learning step of creating a model for extracting an expression having a predetermined attribute for a character feature from newly input extraction sentence data, using the replaced data and the labeled feature data as learning data,
How to create a model. A computer for creating a model for extracting a predetermined expression from text data,
Feature generation means for dividing the learning sentence data into words, and setting feature data for a plurality of features including character features for each word,
In the characteristic data, a labeling unit that labels each word,
Replaced data is generated by replacing, for each expression, one or a plurality of expressions that have a predetermined attribute regarding the character feature included in the learning text data or the feature data, with a randomly generated character string. Replacement means to
Learning means for creating a model for extracting an expression having a predetermined attribute for a character feature from newly input extraction sentence data, using the replaced data and the labeled feature data as learning data,
Program to function as.

Images