JP5322047B2 - Text readability evaluation system - Google Patents

Description

The present invention relates to the ease of evaluation system reading of the sentence.

  As a method for measuring the readability (readability: readability) of a Japanese sentence, for example, there is a function called “evaluation of readability” installed in a Microsoft word. This function is used to (1) the number of characters, (2) the number of words, (3) the number of sentences, (4) the number of paragraphs, (5) the average number of sentences in one paragraph, (6) the average sentence length, (7 8) Analyze into 8 items, the average number of characters between punctuation points, (8) Kanji, Hiragana, Katakana, Alphabet, etc. However, there is no evaluation of how easy or difficult the text is to read.

  Further, Patent Document 1 discloses a method for evaluating the readability of a sentence from the difficulty level of each word in the sentence, but accurate readability cannot be measured only by the difficulty level of each word.

JP 2004-334699 A

The present invention has been made in view of the current situation as described above, and can be used not only as a standard for determining texts for text comprehension research in psychology, Japanese language education, and Japanese language education, but also as an index for text creation. In addition to contributing to the academic field related to sentence comprehension, it can be applied to work to clarify and simplify various reading materials (government publications, warnings about dangerous goods and medicines, business documents, etc.) It is possible to provide useful materials for daily life, and it is easy for not only Japanese people but also foreigners who do not have enough Japanese language learning to share information in Japanese, contributing socially and internationally. degree is intended to provide the ease of evaluation system reading of an extremely high for practical use were Shigeru been sentence.

  The gist of the present invention will be described.

Hiragana for calculating the sentence storage means for storing a target sentence, and the average number-of-characters calculation means for calculating the average number of characters X ₁₁ per sentence sentences this sentence stored in the storage means, the hiragana percentage X ₁₂ of the entire sentence A ratio calculating means; an average predicate number calculating means for calculating an average number of predicates per sentence X ₁₃ of the sentence; an average phrase number calculating means for calculating an average number of phrases X ₁₄ per sentence of the sentence; The average number of characters per sentence X ₁₁ calculated by the average number of characters calculation means, the ratio X ₁₂ of hiragana in the sentence calculated by the hiragana ratio calculation means, the average per sentence calculated by the average predicate number calculation means Using the predicate number X ₁₃ and the average phrase number X ₁₄ per sentence calculated by the average phrase number calculation means as explanatory variables, the following equation (4)
_{Y = a 11 X 11 + b} 11 X 12 + c 11 X 13 + d 11 X 14 + X 00 (4)
However, in the above formula _{(4), a 11, b} 11, c 11, d 11 _{coefficient, X 00}
Comprises an evaluation value deriving means for deriving an evaluation value Y for substituting a constant to evaluate the readability of the sentence as a dependent variable, and the evaluation value deriving means has an evaluation value Y in the first grade to the third grade. Each of the coefficients a ₁₁ , b ₁₁ , c ₁₁ , d ₁₁ and the constant X ₀₀ in the equation (4) is determined so as to determine which of the corresponding level 1 to level 9 is closest. The average number of characters per sentence of each sentence sample obtained by analyzing a large number of sentence samples extracted from Japanese textbooks corresponding to each grade of the first grade to the third grade or books corresponding to the grades X ₁₁ In each of the above grades, which are set in advance in each sentence sample, the ratio X ₁₂ of hiragana in the whole sentence, the average number of predicates per sentence X ₁₃ and the average number of phrases X ₁₄ per sentence in the sentence are explanatory variables. corresponding evaluation value Y Those of the sentence ease evaluation system readings, characterized in that is obtained by performing a multiple regression analysis as the dependent variable.

Since the present invention is configured as described above, it can be used not only as a standard for determining texts for text comprehension research in psychology, Japanese language education, and Japanese language education, but also as an index for text creation. It can be applied to work that makes reading material (government publications, precautions for dangerous materials and medicines, business documents, etc.) clear and easy. It is possible to provide useful materials, and it leads to the sharing of information in Japanese, which is easy for not only Japanese but also foreigners who do not have sufficient Japanese language learning, making a high contribution both socially and internationally. It is extremely easy evaluation system to practical reading of Xiu the sentence.

  The preferred embodiment of the present invention (how to carry out the invention) will be briefly described, showing the operation of the present invention.

Calculate the average number of characters X ₁ per sentence, the average number of words X ₂ per sentence, the ratio of Chinese words X ₃ , the average number of dependency X ₄ per sentence of the target sentence stored in the sentence storage means, A predetermined evaluation value Y is obtained by applying the following formula (1), and the readability of a plurality of target sentences can be objectively compared by this evaluation value Y. In particular, it will be possible to objectively evaluate the readability of Japanese sentences, which was not possible before.

Specifically, for example, the average number of characters X ₁ per sentence of each sentence sample obtained by analyzing a large number of sentence samples, the average number of words X ₂ per sentence, the proportion of Chinese words X ₃ and the average number of dependency Using coefficients a ₁ , b ₁ , c ₁ , d ₁ and a constant X ₀ obtained by performing multiple regression analysis with X ₄ as an explanatory variable and an evaluation value Y preset for each sentence sample as a dependent variable (For example, Japanese textbooks and books with the evaluation value Y as the school year are used as samples.) Furthermore, the evaluation value Y takes a value from 1 to 9, and each is set to correspond to the first grade to the third grade. By doing so, it is possible to intuitively determine the level of the target sentence, and it becomes possible to display the readability of the target sentence objectively and easily.

The average number of characters per sentence X ₁ , the average number of words per sentence X ₂ , the percentage of Chinese words X ₃ , and the average dependency per sentence X ₄ can be easily obtained by existing software. It is possible, and the above multiple regression analysis can be easily performed using existing software. Therefore, easy and accurate evaluation of text readability is realized by combining these easily obtained information. It becomes possible.

Further, as in Example 2 to be described later, the average number of characters X ₁₁ per sentence of the target sentence stored in the sentence storage means, the ratio X ₁₂ of hiragana for the whole sentence, the average number of predicates X ₁₃ per sentence, and When the average number of phrases X ₁₄ per sentence is calculated and given evaluation value Y is obtained by applying it to Equation (4), the readability of a plurality of target sentences can be objectively evaluated more accurately by this evaluation value Y. Can be compared.

  Specific embodiments of the present invention will be described with reference to the drawings.

Example 1, a sentence storage means for storing a target sentence, and the average number-of-characters calculation means for calculating the average number of characters X ₁ per sentence sentence stored in the sentence storage means, the average per one sentence of the sentence Mean word number calculating means for calculating the number of words X ₂ , Chinese word ratio calculating means for calculating the Chinese word ratio X ₃ of the whole sentence, and average relation for calculating the average dependency number X ₄ per sentence of the sentence The received number calculating means, the average number of characters X ₁ per sentence calculated by the average number of characters calculating means, the average number of words X ₂ per sentence calculated by the average word number calculating means, and the Chinese word ratio calculating means Using the calculated Chinese language ratio X ₃ and the average dependency number X ₄ per sentence calculated by the average dependency number calculation means as explanatory variables, the following equation (1)
Y = a ₁ X ₁ + b ₁ X ₂ + c ₁ X ₃ + d ₁ X ₄ + X ₀ (1)
However, in the above formula (1), a ₁ , b ₁ , c ₁ , d ₁ are coefficients, and X ₀ is assigned to a constant, and an evaluation value Y that is used to evaluate the readability of the sentence as a dependent variable is derived. And derivation means.

Specifically, the coefficients a ₁ , b ₁ , c ₁ , d ₁ and the constant X ₀ in the formula (1) are averages per sentence of each sentence sample obtained by analyzing a large number of sentence samples. Multiple regression analysis is performed using the number of characters X ₁ , the average number of words per sentence X ₂ , the ratio of Chinese words X ₃ and the average dependency number X ₄ as explanatory variables, and the evaluation value Y preset for each sentence sample as the dependent variable. It is obtained by doing.

  In the first embodiment, the readability of a sentence is measured by four factors: the length of a sentence by the number of characters, the length of a sentence by the number of words, the ratio of word types, and the complexity of the grammar. Shows up to levels. As shown in Table 1, Levels 1 to 9 represent first grade to third grade in school education.

  For sampling of basic data to determine the above coefficients and constants, from the first grade to the sixth grade, Mitsumura Publishing, school books, and 36 books of the first and second volumes of Tokyo books, a total of 36 books, junior high school grades 1 to 3 Until the year, a total of 45 books, including elementary school textbooks and junior high school textbooks, were used, with a total of nine books: Mitsumura Publishing, Sanseido, and Tokyo Books.

  Materials intended for reading from each textbook were targeted, but those intended for learning characters, vocabulary, writing, discussion, and presentation were excluded. In addition, poetry is excluded from the subject because it differs greatly from prose in form and expression. As a result, 264 texts were selected from 45 textbooks (Table 2).

  In addition, from these, duplicate texts (first grade "Okina Kabu", second grade "Kasako Jizou" "Swimmy", third grade "Mochimochi no Ki", fourth grade "White Boshi" "Pole Pole" “Gongitsune,” “One Flower,” 5th grade elementary school, “Old Taijisan and Cancer,” “Highly Ordered Restaurant,” 2nd grade, “Runos Meros,” 3rd grade, “Hometown,” etc. are listed in multiple textbooks. And 245 texts were used as data in the end, except for special texts (eg, “Jugemu” in the third grade of elementary school has an extremely large number of characters per word).

  Each explanatory variable will be described below.

  The average number of characters per sentence will be described. In the field of memory research, it has become clear that the longer the sentence, the harder it is to read and the shorter the sentence, the easier it is to read. The length of a sentence can be expressed by the number of characters per sentence. First of all, we put all the texts to be analyzed on a scanner page by page, took images with eTypistv11.0 (software), and made them available on the computer screen.

  Next, the average number of characters per sentence was calculated for all texts using “readability assessment”, one of Microsoft's word functions. When the texts are summarized for each school year, it can be seen that the average number of characters per sentence increases with each school year, as shown in FIG.

  The average number of words per sentence will be described. In a language with only one alphabet, such as English, the average number of characters per sentence reflects the length of the sentence as it is. However, there are four types of Japanese characters, kanji, hiragana, katakana, and romaji (alphabet), so I thought it was not appropriate to indicate the length of a sentence only by the number of characters. For example, the sentence (1) “The liberal arts course will be opened” is 10 characters, while the sentence (2) “You are being asked” is 16 characters.

  Actually, a sentence like (2) is not used in daily life, but the amount of kanji used is left to the writer's judgment and there is no standard. In order to solve this problem, the present inventors considered as follows.

  (1) and (2) differ in the number of characters, but the number of words is the same, and the amount of information contained in the sentence is also the same. For example, (1) “Education / Course / Guide / Opened” and (2) “Kyoyo / Kate / Gai / Kaisei” are sentences composed of four words. In this way, the length of a sentence can be indicated by the number of words, not only in Japanese but also in a language in which a plurality of character types exist. Therefore, we used the Microsoft word function to calculate the average number of words per sentence for each text. Summarizing the text by grade, it can be seen that the average number of words per sentence increases with each grade, as shown in FIG.

  Explain the ratio of word types. The Japanese vocabulary includes a plurality of word types such as Chinese, Japanese, foreign words, mixed words, and in Example 1, the ratio of the word types is one of the determinants of the readability of the sentences. The following method was used to calculate the percentage of word types.

  First, morphological analysis was performed on the entire text using ChaSen (software), and the number of four types of content words: nouns, verbs, adjectives, and adjective verbs was calculated. Next, using Katarigusa (software), the content words are analyzed into Chinese, Japanese, foreign words, mixed languages, Chinese foreign languages, Japanese foreign languages, and Japanese-Korean languages. The ratio of each mixed word was calculated. As a result, it was shown that the ratio of foreign words and mixed words did not change according to grades, while the sum of the percentages of Japanese and Chinese in each grade was almost constant. In other words, as the school year goes up, the Japanese language decreases, and conversely the Chinese language increases. Fig. 2 shows the percentage of word types by grade.

  Explain the complexity of grammar. In the first embodiment, the complexity of the grammar will be described with the number of dependent clauses and the distance between the clauses. Each sentence of all texts is analyzed with Japanese dependency analyzer CaboCha (software), the total number of dependency for each text, and the distance between the dependency clause and the reception clause (between the dependency clause and the reception clause). The number of phrases) was calculated. In the analysis result in CaboCha, the maximum number of clauses between the dependency clause and the reception clause is 35, but it is difficult to think that there are actually 35 phrases between the dependency clauses. This seems to be a problem caused by the accuracy of CaboCha being 89.29%. Therefore, in Example 1, only the case where the number of clauses is 1 to 10 was set as the analysis target.

  Table 3 shows, from the left, the text name, the total number of dependency in the text, and the number of clauses between the dependency clause and the reception clause. The distance between the dependency clause and the receiving clause indicates 1 when the receiving clause is immediately adjacent to the dependency clause, and 2 indicates the case when there is one clause between the dependency clause and the receiving clause. is there. For example, text1 is a text of a first grader of elementary school, but the total number of dependency is 32, and the one whose distance is 1 (the one with the reception clause right next to the dependency clause) is 16, which is 50% of the total. And since there are almost no distances greater than 4, it can be seen that text1 is a simple sentence when viewed from the amount and complexity of the dependency. On the other hand, text4 is a text of junior high school third grader, but out of the total number of dependency 410, the one with distance 1 is 66, which is 16% of the total, and there are many distances from 2 to 10 with complicated sentences. I know that there is.

Therefore, in this study, the following formula (2) was used to make the grammar complex.

  Using the above formula, the complexity of the grammar (number of dependencies per sentence) is calculated for each grade, and as shown in Table 4 and Figure 3, the complexity increases as the grade rises. It was shown that.

  From the above, the four factors, the average number of characters per sentence, the average number of words per sentence, the proportion of Chinese, and the complexity of the grammar (number of dependency per sentence), are all numerical values as the grade increases. Since it was shown that the grade would be linear, the average number of characters per sentence, the average number of words per sentence, Chinese, from the 245 texts in the textbooks from the first grade to the third grade, assuming the grade was linear The numerical values of the ratio and the complexity of the grammar were calculated, and these four variables were used as explanatory variables, and the grade was used as the dependent variable, and analyzed by multiple regression analysis (forced input method) using statistical analysis software SPSS.

  As a result, it was shown that 22 texts deviated greatly from the predicted values. Therefore, these 22 texts were excluded as outliers, and finally multiple regression analysis was performed again with the remaining 223 texts. The result is shown in FIG.

  The part indicated by β is the coefficient. Since all four variables are significant (p <.001), it is shown that it is appropriate to use these four variables as explanatory variables. When the coefficient shown here is applied to each variable, the following equation (3) is obtained.

_{Y = 0.171X 1 -0.439X 2 + 13.434X} 3 + 0.065X 4 +2.388 (3)
Y: Grade, X ₁ : Average number of characters per sentence, X ₂ : Average number of words per sentence, X ₃ : Percentage of Chinese, X ₄ : Average number of dependency per sentence

  The R-square value is a coefficient of determination indicating the strength of the predictive power, but 0.668 is a high value (1.0 is the highest value), so the predictive power of this equation is high. The grade indicated by Y is a judgment value of readability, and takes values from 1 to 9. 1 is the simplest and 9 is the most difficult. If four variables of the text are calculated and applied to this formula, it can be determined which grade of the textbook the text is closest to. Therefore, this is a readability formula that predicts the grade level of the text. With this formula, anyone can develop software that calculates each variable and anyone can easily determine the level of text.

  Each means of the first embodiment will be described.

  The first embodiment can be implemented using a general computer (PC or the like). Specifically, the text storage means is a storage device such as a hard disk, an optical disk, or a memory, and an arithmetic processing unit such as a CPU functions as the following means based on a program stored in the storage device such as the hard disk. It is configured as follows.

The average number-of-characters calculation means reads the target sentence that is stored in the text storage means, the average number of characters X ₁ per sentence of the target sentence is for calculating the number of a whole text characters ÷ statement, the calculation result Store in the storage device.

Mean word number calculating means reads the target sentence that is stored in the text storage means, the average number of words X ₂ per sentence of the target sentence, after divided into each word of a whole text, number of words in a whole text ÷ It is calculated by the number of sentences, and the calculation result is stored in the storage device.

Chinese ratio calculation means reads the target sentence that is stored in the text storage means, the Chinese ratio X ₃ of the entire target sentence, after dividing the entire sentence to the Chinese and other Katarigusa, the entire text of Chinese It is calculated by the number divided by the content word of the whole sentence, and the calculation result is stored in the storage device.

The average dependency number calculating means reads the target sentence stored in the sentence storage means, and calculates the dependency number X ₄ per sentence of the target sentence using the above equation (2). The result is stored in the storage device.

The evaluation value deriving means calculates the evaluation value Y by reading out X _{1 to} X ₄ calculated by each of the above means and stored in the storage device and substituting it into the above equation (3). Either stored in a storage device or displayed on an appropriate display means such as a display (X _{1 to} X ₄ and the process may be displayed together).

Therefore, for example, a program serving as each of the above means is stored in a storage device such as a PC, and the coefficients a ₁ , b ₁ , c ₁ , d ₁ and constants in the above equation (1) are stored using appropriate samples. to obtain X ₀ in advance, by providing the evaluation value deriving means an expression like the equation (3) or the like, inputs the a X ₁ to X ₄ of the target sentence to easily analyze the existing software or the like You can evaluate the readability of Japanese sentences just by doing.

  Therefore, in Example 1, not only a standard for determining texts for text comprehension research in psychology, Japanese language education, and Japanese language education, but also an index for text creation, various daily readings (government publications, It can be applied to work to clarify and simplify dangerous goods and chemicals, business documents, etc.), and can provide materials that are useful not only for academic fields related to sentence comprehension but also for everyday life of the general public. Moreover, it is easy to share information in Japanese, not only for Japanese but also for foreigners who are not good at learning Japanese, and it is extremely practical because it contributes both socially and internationally. It will be.

In the second embodiment, a part of the formula of the first embodiment is improved to enable more accurate evaluation. Specifically, a sentence storage means for storing a target sentence, and the average number-of-characters calculation means for calculating the average number of characters X ₁₁ per sentence sentences this sentence stored in the storage unit, the entire sentence Hiragana proportion X Hiragana ratio calculating means for calculating the _12, average clause number for calculating the average predicate count calculation means for calculating the average predicate number X ₁₃ per sentence, the average clause number X ₁₄ per sentence of the sentence of the sentence An average number of characters X ₁₁ per sentence calculated by the calculating means, the average number of characters calculating means, a ratio of hiragana X ₁₂ in the sentence calculated by the hiragana ratio calculating means, and an average predicate number calculating means Using the average predicate number X ₁₃ per sentence and the average phrase number X ₁₄ per sentence calculated by the average phrase number calculating means as explanatory variables, the following equation (4)
_{Y = a 11 X 11 + b} 11 X 12 + c 11 X 13 + d 11 X 14 + X 00 (4)
However, in the above formula _{(4), a 11, b} 11, c 11, d 11 _{coefficient, X 00}
Is provided with an evaluation value deriving means for deriving an evaluation value Y, which is substituted into a constant, to evaluate the readability of a sentence as a dependent variable.

  In the formula used in Example 1, the average number of words per sentence is one of the variables. However, after the Japanese language textbook is converted into an electronic file, it is analyzed using Microsoft's Word and morphological analysis tool ChaSen. As a result of detailed examination, we found that there are some problems in how to count compound words. For example, Nagaoka University of Technology is a proper noun with one word, but the Word function counts as four words, such as “Nagaoka / Technology / Science / University”. This is because there is no Nagaoka University of Technology in the thesaurus (like a dictionary) that is the basis of Word. ChaSen is a tool originally created to analyze newspaper articles, so it is not suitable for sentences with many hiragana characters, and there is a drawback in the analysis results of compound words. Japanese has a rule that phonology changes when a compound word is formed by a combination of a plurality of words. For example, the sounds of “Hako” and “Hatake” change, such as “Honba” and “Hana” between “Hon” and “Hako”, “Hana” and “Hatake”. Common vocabularies such as “Book Box” and “Hanabata” are in the ChaSen thesaurus, but there are no uncommon compound words. Therefore, when analyzing the third grade elementary school language teaching material “Gongitsune”, one word “mother fox” is analyzed as “mother (noun) gi (unknown word) tsune (person name, proper noun)”. This is because “Mother”, “Kitsune”, and “Tatsumi” have no “Mother's fox” even in the ChaSen thesaurus. A similar problem is found not only in such compound words but also in onomatopoeia, and “tonton” is analyzed like “ton” (person name, proper noun). A part of the analysis results is shown in Table 5.

  From the above, it is considered that the reliability of the average number of words per sentence by the analysis tool is somewhat insufficient, and in Example 2, the average number of words per sentence is excluded from the official variables. .

  In addition, the formula used in Example 1 reflects the complexity of the grammatical structure, and the average number of dependency per sentence is one of the variables. However, there is a problem with this point as well. discovered.

  Japanese grammar has a nested structure (a structure in which multiple phrases can be placed in the subject-predicate), and the complexity of the grammar structure can be explained only by the distance and number of the related and receiving clauses. It may be insufficient. For example, in the following 1), there are two dependencies: “The boy throws” and “The boy runs away”. 2) But there are two dependencies: "The boy threw" and "The ball fell". However, in 1) “throw” and “run away” have a parallel structure relationship, whereas in 2) “the boy throws” and “ball” have a combination modification relationship, I thought it was impossible. For this reason, it was considered that the average dependency per sentence in the official variables represents the complexity of the grammar, so we decided to exclude it.

1) The boy threw the ball and ran away.
2) The ball thrown by the boy fell into the river.

  Based on the above points, exclude the average number of words per sentence and the average number of dependency per sentence from the variables that predict the grade, and instead of the average number of words per sentence, the average number of phrases per sentence I decided to use Also, instead of the average number of dependency per sentence, we decided to use the average number of predicates per sentence. The reason is described below.

  The reason why the average number of words per sentence is included in the formula used in Example 1 is that there are a plurality of character types in Japanese, and the average number of characters per sentence is not enough to indicate the length of the sentence. Because I thought it was enough. If the average number of words per sentence cannot be used to indicate the length of a sentence, the number of clauses per sentence can be used as a variable indicating the length of the sentence as an alternative. According to Hashimoto Grammar (School Grammar) by Hashimoto Shinkichi, a phrase is “in Japanese, a unit of pronunciation that is connected to an independent word (noun, verb, etc.). It is said that it would not be strange to put “ne” or “sa” in the middle of the sentence. For example, the sentence “The ball thrown by the boy fell into the river” can be divided into “the boy / thrown / the ball / falled into the river.” It can be said.

  In the field of language processing, clauses can be counted with the dependency analysis tool CaboCha. CaboCha also basically analyzes the dependency based on the morphological analysis result of ChaSen, but if it is a unit of a phrase, a particle comes immediately after it even if the word is a compound word. There are few mistakes such as disassembling and counting. In addition, even if there is, it is easy to find artificially. FIG. 5 shows the calculation of the average number of phrases per sentence for each grade from 1 (1st grade of elementary school) to 12 (3rd grade of high school). As shown in FIG. 6, it rises almost linearly according to the school year, and is considered suitable as a variable. Therefore, I decided to use this as a variable.

  A predicate is a component that plays a central role in sentences and clauses in linguistics, and represents a part related to other noun phrases. A statement with one predicate is called a simple statement, and a statement with two or more predicates is called a compound statement. Because compound sentences are more complex than simple sentences, it is more difficult for readers to have more than one sentence. In addition, an experiment by Goetz, Andersion, & Schallert conducted in 1981 shows that “the more propositions are more difficult if the sentence length is the same”. Since the proposition here is a unit consisting of one predicate and one term, the number of propositions is obtained by counting the predicates. Therefore, it is considered that the complexity of the grammatical structure, which is difficult to explain with the number of dependencies, can be explained to some extent with the number of predicates per sentence. The average number of predicates per sentence for each grade is shown in FIG.

  In Example 1, Japanese language textbooks of grades 1 (first year of elementary school) to 9 (third year of middle school) were used as samples. Nineth grade seems to be sufficient to create a formula for predicting the school year, but in Example 2, the data from the first to third year of high school was further increased and observed for 12th grade. This is because it is possible to confirm whether the grade is “linear” by increasing the grade. However, as a result, it was shown that up to the ninth grade was “linear”, so only the data for the ninth grade was used for the analysis.

  Furthermore, considering that there are four types of characters (Hiragana, Kanji, Katakana, and Roman) as one of the major features that make Japanese different from other languages, the proportion of character types is also a strong variable in the readability formula. It is expected that. In Example 1, this point was not included, but in Example 2, the ratio of the character type was also added as a variable. The types of characters appearing in Japanese language textbooks from the first grade to the third grade were calculated for each grade and analyzed as shown in FIG.

  Hiragana decreases with each grade, and kanji increases with each grade. Either one can be used as a variable. However, many texts in the first grade of elementary school have zero occurrences of kanji. Since it is expected that including a sample indicating zero may be inappropriate in the calculation of statistical analysis, the ratio of hiragana was used as a variable.

  Based on the above analysis, the variables for constructing the readability formula include the average number of characters per sentence and the percentage of Chinese used in Example 1, the average number of phrases per sentence, the percentage of hiragana and 1 Five variables of the average number of predicates in the sentence were judged to be appropriate as independent variables for predicting grade.

  Observing changes in each variable according to grades, it is almost right up to grade 9, so it seems safe to assume that grades 1 to 9 are linear. (See FIGS. 6-9). Although the text analyzed in Example 2 and the text analyzed in Example 1 partially overlap, they are not exactly the same.

  Therefore, using the data from the first grade to the ninth grade from all data, the above five variables are made independent variables, and the grade is made a dependent variable. Forced input method). The forced input method is an analysis method that uses all variables with strong explanatory power and weak variables as effective variables. As a result, 40 texts deviating from the predicted values were excluded. In Example 1, 22 texts were excluded, but in Example 2, the outlier criteria were tightened and 40 texts were excluded in order to achieve higher accuracy. Using only the remaining 205 texts, as in the first method, a multiple regression analysis was performed with the grade as a dependent variable and the other variables as independent variables (stepwise method). The stepwise method is an analysis method that can leave only explanatory variables. Despite the fact that Kanji has a strong correlation with school year (r = 0.749, p <.0001), it was excluded because of the extremely high correlation between the ratio of Kanji and hiragana (r = -0.870, p <.0001). ), Probably due to multicollinearity. Multicollinearity means that two or more variables explain the same thing. Using the stepwise method, only one variable explaining the same thing is left, and the rest is excluded. It has become. Table 6 shows the average value and standard deviation for each variable including the Chinese language and the correlation between the variables.

  FIG. 10 shows the results of multiple regression analysis for four variables excluding the proportion of Chinese. For constants and coefficients see Model 4 non-standardized coefficients. The percentage of hiragana in the sentence is -0.148, the average number of predicates in one sentence is 1.585, the average number of characters in a sentence is -0.117, and the average number of phrases in a sentence is -0.126, which are coefficients. The constant is 15.581. Although not shown in FIG. 10, the R-square value was 0.858. The R-square value represents the certainty of the predictive power of the formula calculated by the multiple regression analysis, and is shown between the lowest value zero and the highest value 1.0. The number 0.858 indicates that this formula has a very strong predictive power in predicting grades.

  When the presence or absence of multicollinearity between four variables was diagnosed, in the model with eigenvalue of 0.008 and condition index of 24.206, the variance ratio was 0.45 average predicate number per sentence, 0.99 average number of characters per sentence, and 0.42 average phrase number per sentence. The result showed that there was a suspicion of collinearity between these three variables. Therefore, in order to be cautious, the number of predicates is excluded from this, and multiple regression analysis (stepwise method) is performed with three variables: the average number of characters in one sentence, the ratio of hiragana and the average number of phrases in one sentence. As a result, the average number of characters in one sentence and the average number of phrases in one sentence were excluded, and only the ratio of hiragana remained. However, even in the mathematical calculation, it is practically unnatural that only the ratio of hiragana is a variable that predicts the school year. In addition, as described above, since the R square value as an explanatory variable in the four variables of the average number of characters in one sentence, the ratio of hiragana, the average number of predicates in one sentence, and the average number of clauses in one sentence is 0.858, it is extremely high. The following formula (5) was created by judging that the formula using these four variables was valid, multiplying each variable by a constant, and adding a constant.

_{Y = -0.117X 11 -0.148X 12 + 1.585X} 13 -0.126X 14 +15.581 (5)
Y: Grade, X ₁₁ : Average number of characters per sentence, X ₁₂ : Ratio of hiragana characters in the whole sentence,
X ₁₃ : average number of predicates per sentence, X ₁₄ : average number of clauses per sentence

  Therefore, if the four variables of the text are calculated and applied to the above formula (5), it is possible to determine more precisely which grade of the textbook in the Japanese language textbook is closer to the case where the formula according to the first embodiment is used.

  Each means of Example 2 will be described.

  The average number-of-characters calculation means is the same as in the first embodiment.

Furthermore, Hiragana ratio calculation means reads the target sentence that is stored in the text storage means, the hiragana percentage X ₁₂ in the overall target sentence, after dividing the entire sentence to the hiragana and other types of characters, the whole sentence It is calculated by the number of hiragana divided by the number of characters in the entire sentence, and the calculation result is stored in a storage device.

The average predicate count calculation means reads the target sentence that is stored in the text storage means is for calculating a predicate number X ₁₃ per sentence of the target sentence, one that stores the calculation result in the storage device is there.

The average phrase number calculation means reads the target sentence stored in the sentence storage means, calculates the number of phrases X ₁₄ per sentence of the target sentence, and stores the calculation result in the storage device. is there.

The evaluation value deriving means calculates the evaluation value Y by reading X _{11 to} X ₁₄ calculated by each of the means and stored in the storage device and substituting it into the above equation (5). Either stored in the storage device or displayed on an appropriate display means such as a display (X _{11 to} X ₁₄ and the process may be displayed together).

Therefore, for example, a program serving as each of the above means is stored in a storage device such as a PC, and the coefficients a ₁₁ , b ₁₁ , c ₁₁ , d ₁₁ and constants in the above equation (4) are stored using appropriate samples. obtained in advance the X _00, by providing the evaluation value deriving means an expression like the equation (5) or the like, enter this X ₁₁ to X ₁₄ of the target sentence to easily analyze the existing software or the like You can evaluate the readability of Japanese sentences just by doing.

  The rest is the same as in Example 1.

  In addition, Example 2 has confirmed that the readability of a Japanese sentence can be evaluated appropriately as the following comparative experiments.

  On October 10th, 2007, Sato Lab at Nagoya University published a page called “Language Box” and a tool called “Measure the difficulty of Japanese text”. 18 texts were selected at random from the teaching materials actually published in the textbooks of Japanese language. When the school year was calculated using both this tool (hereinafter referred to as a comparative example) and the readability formula of Example 2 (hereinafter referred to as an example), the results shown in Table 7 were obtained. Analysis of variance was performed for the actual grade of the text, the grades obtained in the examples and comparative examples by repeated measurements. As a result, significant main effects were obtained for the three grades: actual grade (M = 5.28, SD = 2.47), examples (M = 5.11, SD = 2.40), and comparative examples (M = 3.78, SD = 2.58). [F (2,34) = 26.404, p <.001]. This means that there is a difference between the three. Then, a simple comparison was made to see where there was a difference. As a result, there was no significant difference between the actual grade and the grade index of the working example [F (1,17) = 1.889, p = .187, ns], but there was a significant difference between the actual grade and the comparative example. [F (1,17) = 33.585, p <.001]. In other words, there is no significant difference between the results calculated in the actual school year and the example, and if any, there is a slight difference, but there is a difference between the results calculated in the actual school year and the comparative example. Furthermore, a significant difference was also observed between Examples and Comparative Examples [F (1,17) = 24.727, p <.001]. Therefore, it was shown that the example can predict the actual grade of a text more appropriately than a comparative example.

  The present invention is not limited to this embodiment, and the specific configuration of each component can be designed as appropriate.

It is a graph which shows the average number of letters and average number of words per sentence for every grade. It is a graph which shows the ratio of the word type for every grade. It is a graph showing the complexity of grammar for each grade. It is explanatory drawing which shows the result of a multiple regression analysis. It is a graph which shows the average number of clauses per sentence for every grade. It is a graph which shows the average number of predicates per sentence for every grade. It is a graph which shows the ratio of the character type for every grade. It is a graph which shows the number of characters per sentence and the number of clauses for every grade. It is a graph which shows the ratio of the word type for every grade. It is explanatory drawing which shows the result of a multiple regression analysis.

Claims (1)

Hiragana for calculating the sentence storage means for storing a target sentence, and the average number-of-characters calculation means for calculating the average number of characters X ₁₁ per sentence sentences this sentence stored in the storage means, the hiragana percentage X ₁₂ of the entire sentence A ratio calculating means; an average predicate number calculating means for calculating an average number of predicates per sentence X ₁₃ of the sentence; an average phrase number calculating means for calculating an average number of phrases X ₁₄ per sentence of the sentence; The average number of characters per sentence X ₁₁ calculated by the average number of characters calculation means, the ratio X ₁₂ of hiragana in the sentence calculated by the hiragana ratio calculation means, the average per sentence calculated by the average predicate number calculation means Using the predicate number X ₁₃ and the average phrase number X ₁₄ per sentence calculated by the average phrase number calculation means as explanatory variables, the following equation (4)
_{Y = a 11 X 11 + b} 11 X 12 + c 11 X 13 + d 11 X 14 + X 00 (4)
However, in the above formula _{(4), a 11, b} 11, c 11, d 11 _{coefficient, X 00}
Comprises an evaluation value deriving means for deriving an evaluation value Y for substituting a constant to evaluate the readability of the sentence as a dependent variable, and the evaluation value deriving means has an evaluation value Y in the first grade to the third grade. Each of the coefficients a ₁₁ , b ₁₁ , c ₁₁ , d ₁₁ and the constant X ₀₀ in the equation (4) is determined so as to determine which of the corresponding level 1 to level 9 is closest. The average number of characters per sentence of each sentence sample obtained by analyzing a large number of sentence samples extracted from Japanese textbooks corresponding to each grade of the first grade to the third grade or books corresponding to the grades X ₁₁ In each of the above grades, which are set in advance in each sentence sample, the ratio X ₁₂ of hiragana in the whole sentence, the average number of predicates per sentence X ₁₃ and the average number of phrases X ₁₄ per sentence in the sentence are explanatory variables. corresponding evaluation value Y Sentence ease evaluation system readings, characterized in that is obtained by performing a multiple regression analysis as the dependent variable.

Images