JP6502279B2 - Outlier location extraction device, method and program - Google Patents

Description

  The present invention relates to a technique for determining and / or extracting a fragment out of at least one fragment included in a text whose degree of difficulty is higher than a predetermined standard.

  Research to estimate the degree of difficulty of the text has been conducted since ancient times. However, many of them have been proposed for the purpose of estimating the degree of difficulty of the entire text, and can not be used to support writing by feeding back information on some sentences, phrases, and phrases.

  For example, in Non-Patent Document 1, only the character bigram is used as a feature amount for estimating the difficulty level. Therefore, the target text is required to include a certain amount of characters or more. Further, Non-Patent Document 1 reports that if the effective character bigram is 25 or more, the degree of difficulty can be estimated with a high correlation of 0.9 or more (see, for example, Non-Patent Document 1).

  Further, Non-Patent Document 2 proposes a multiple regression equation for estimating a target grade (degree of difficulty) using the average number of predicates contained in one sentence and the proportion of hiragana in the entire text as variables. In the case of Non-Patent Document 2, there is an advantage that the amount of text required for estimation may be small. However, since only the number and ratio of predicates are used, it is not possible to feed back the difficulty of each word or the like (see, for example, Non-Patent Document 2).

Kengo Kojima, et al., "Estimated difficulty of Japanese text using a character bigram model", Proceedings of the Language Processing Society of Japan, The 15th Annual Conference, Proceedings, March 2009, pp. 897-900 SHINOZAKI Hideko, 1 other person, "Construction of writing difficulty grade grade for elementary and junior high schools based on Japanese language textbooks", Journal of Japan Society for Educational Technology, Vol. 33, No. 4, pp. 449-458, 2010

  The previous difficulty level estimation techniques have been proposed for the purpose of estimating the level of difficulty of the entire text, and show that the level of difficulty required for a part of the text does not match or does not match Was unable to give feedback.

  An object of the present invention is to determine a fragment having a degree of difficulty exceeding a predetermined level among at least one fragment included in the text as an outlier point of the text and / or as an outlier point of the text An outlier location extracting apparatus, method and program for extracting are provided.

The outlier location extracting device according to one aspect of the present invention includes a fragment difficulty level estimation unit for estimating the difficulty level class of each fragment obtained by dividing the text in a predetermined unit, included in the input text; Have a degree of difficulty away from the estimated class of difficulty of the text, based on a comparison of the overall class of difficulty estimation part and the class of difficulty of each estimated segment and the class of estimated complexity of the text And an outlier location extraction unit for extracting an outlier location which is a text fragment, and the fragment difficulty level estimation unit is configured to calculate the appearance frequency of each word n-gram included in each segment and the difficulty level for each class. The likelihood of each fragment belonging to each difficulty class is estimated based on the occurrence probability of each word obtained in advance, and the difficulty class of the highest likelihood is defined as the difficulty class of each fragment, Overall difficulty estimation part Estimate the likelihood that the text belongs to each difficulty level based on the appearance frequency of each word n-gram included in the text and the occurrence probability of each word previously determined for each difficulty level class The degree of difficulty is assumed to be the estimated degree of difficulty of the text .

The outlier location extracting device according to one aspect of the present invention includes a fragment difficulty level estimation unit for estimating the difficulty level class of each fragment obtained by dividing the text in predetermined units, included in the input text, and each estimated fragment And an outlier location extraction unit for extracting an outlier location, which is a fragment of text, having a degree of difficulty separated from the predetermined difficulty level class based on comparison between the difficulty level class and the predetermined difficulty level class The fragment difficulty level estimation unit determines the difficulty level of each fragment based on the appearance frequency of each word n-gram included in each fragment and the occurrence probability of each word previously determined for each difficulty level class. The likelihood belonging to the class is estimated, and the highest likelihood difficulty class is taken as the difficulty class of each fragment estimated .

  Among the at least one fragment included in the text, a fragment whose degree of difficulty is higher than a predetermined level can be determined to be an outlier point of the text and / or extracted as an outlier point of the text.

The block diagram for demonstrating the example of an outlier location extraction apparatus. The block diagram for demonstrating the example of an outlier location extraction apparatus. The flowchart for demonstrating the example of the outlier location extraction method. The figure which shows the example of a first appearance difficulty level class. The figure which shows the example of the closeness degree of a fragment. The figure which shows the example of the closeness degree corresponding to a difficulty level. The figure which shows the example of the average word number in each difficulty level class. The figure which shows the example of the occurrence probability of a word. The figure which shows the example of the occurrence probability of a word.

  Hereinafter, embodiments of the outlier location extraction apparatus and method will be described with reference to the drawings.

First Embodiment
As shown in FIG. 1, the outlier location extracting apparatus according to the first embodiment includes a preprocessing unit 1, a fragment difficulty level estimating unit 31, an overall difficulty level estimating unit 32, an outlier location extracting unit 4, and an alternative expression presenting unit 5. For example, the alternative expression replacing unit 6, the storage unit 71, the storage unit 72, and the alternative expression storage unit 8 are provided.

  The outlier location extraction method is realized, for example, by each unit of the outlier location extraction apparatus performing the process from step S1 to step S6 in FIG. 3.

<Pre-processing unit 1>
Text is input to the preprocessing unit 1.

  The preprocessing unit 1 performs preprocessing such as morphological analysis, syntax analysis, eigenexpression extraction, and term structure analysis on the input text, and the result of the preprocessing is the fragment difficulty level estimation unit 31 and the overall difficulty level It outputs to the estimation part 32 (step S1). When morphological analysis is performed as preprocessing, the result of morphological analysis is output.

  An existing technique can be used as a pre-processing technique such as morphological analysis, syntactic analysis, specific expression extraction, and term structure analysis.

Hereinafter, an example of a result of morphological analysis which is an example of preprocessing will be described. The following is an example of the result of morphological analysis of the text "To be a trigger for parent-child communication".
"Parent": noun, common noun, general, *, *, *, oyster, parent and child, parent and child, oyster, parent and child, oyster, sum, *, *, *, *
"No": particle, case particle, *, *, *, *, no, of, no, of, no, sum, *, *, *, *
"Communication": noun, common noun, changeable, *, *, *, communication, communication-communication, communication, communication, communication, communication, outside, *, *, *, *
"No": particle, case particle, *, *, *, *, no, of, no, of, no, sum, *, *, *, *
"Occurrence": noun, common noun, general, *, *, *, keiki, opportunity, occasion, cake, opportunity, cake, han, *, *, *, *
"Ni": particle, case particle, *, *, *, *, ni, to, ni, to, ni, sum, *, *, *, *
"Naru": Verb, non-independent, *, *, 5 steps-La line, adjective form-general, null, consists, becomes, null, becomes, null, sum, *, *, *, *
"Yo": shape verb, auxiliary verb stem, *, *, *, *, yo, yo, yo, yo, yo, yo, yo,, *, *, *, *
"Ni": auxiliary verbs, *, *, *, auxiliary verbs-da, continuous forms-da, da, d, di, da, da, sum, *, *, *, *

<Shard difficulty level estimation unit 31>
The fragment difficulty level estimation unit 31 receives the text and the result of preprocessing.

  The fragment difficulty level estimation unit 31 estimates the difficulty level class of each fragment included in the input text (step S31). The degree of difficulty class of each fragment estimated is output to the outlier location extraction unit 4.

  The input text is composed of a plurality of fragments divided in predetermined units. The predetermined unit is an element constituting text such as at least one word, word n-gram, at least one sentence or the like. In other words, a fragment is an element that constitutes text such as at least one word, word n-gram, at least one sentence, and the like. As an example of word n-gram, it is word unigram which is word n-gram in case of n = 1, word n-gram which is word n-gram in case of n = 2, word n-gram in case of n = 3 The word trigram can be mentioned.

  Since a paragraph is composed of at least one sentence, when at least one sentence is a fragment, the case where each paragraph of text is a fragment is included.

  For example, the feature quantity extraction unit 111 in the fragment difficulty level estimation unit 31 extracts the feature quantity of each fragment, and the fragment difficulty level estimation unit 31 uses the feature quantity of each fragment extracted and the difficulty class of the fragment presume. For example, the difficulty level class corresponding to the feature quantity of each estimated fragment is taken as the difficulty level class of each fragment. The feature quantity of the fragment may be any feature quantity as far as the difficulty level of the fragment is concerned. For example, when the fragment is at least one sentence, the feature quantity of the fragment is defined as the average number of words constituting the fragment, the number of phrases constituting the fragment, the ratio of kanji in the fragment, the ratio of katakana in the fragment The ratio of hiragana in the fragment, the ratio of kanji and katakana in the fragment, the ratio of active or passive in the fragment, or the depth of the syntactic tree in the fragment in the fragment can be used.

  The ratio of kanji in the fragment is, for example, (number of kanji characters in the fragment) / (number of characters of the fragment).

  The ratio of katakana in the fragment is, for example, (number of katakana characters in the fragment) / (number of characters of the fragment).

  The percentage of Hiragana in the fragment is, for example, (number of katakana characters in the fragment) / (number of characters of the fragment).

  The ratio of kanji and katakana in the fragment is, for example, (number of kanji and katakana characters in the fragment) / (number of characters of the fragment).

  The ratio of active or passive in the fragment is, for example, (number of appearances of active or passive in the fragment) / (number of appearances of verb in the fragment).

  When the fragment is a word n-gram, the fragment difficulty level estimation unit 31 refers to the information of the association between each fragment and the difficulty class of each fragment, which is stored in advance in the storage unit 71. You may then ask for the difficulty class of each piece. For example, the difficulty class in which a certain fragment first appears, in other words, the lowest difficulty class in which the fragment appears, is stored in the storage unit 71 as the difficulty of the fragment. An example of the first appearance difficulty level, which is a difficulty level in which a certain fragment first appears, is shown in FIG. The fragment difficulty level estimation unit 31 can obtain the difficulty level class of each fragment by reading the difficulty level class corresponding to each fragment from the storage unit 71.

  Alternatively, for example, a difficulty class in which a certain fragment frequently appears, in other words, a difficulty class in which the fragment appears better, may be stored in the storage unit 71 as the difficulty of the fragment.

<Overall difficulty level estimation unit 32>
Text and the result of preprocessing are input to the overall difficulty level estimation unit 32.

  The overall difficulty level estimation unit 32 estimates the difficulty level class of the entire input text (step S32). For simplicity of description, the difficulty class of the entire text is also referred to simply as the "text difficulty class".

  When the feature amount of each fragment is extracted in the fragment difficulty level estimation unit 31, the overall difficulty level estimation unit 3 uses the feature amount of each extracted fragment, and the difficulty level class of the input text Estimate For example, the overall difficulty level estimation unit 3 may calculate the average value of the feature amounts of the fragments included in the input text and set the average as the difficulty level of the input text. Also, for example, when the average value of the sentence length is calculated as the feature amount, the difficulty level class may be obtained by performing threshold determination on the average value of the sentence length. As a result, it is possible to set a difficulty level class that reflects the tendency that the degree of difficulty is high in a text having many long sentences.

  Further, the feature extraction unit 321 in the total fragment difficulty estimation unit 32 may obtain the feature of the text, and the entire difficulty estimation unit 32 may use the obtained feature as the text difficulty class. The feature quantity of the text may be any feature quantity related to the degree of difficulty of the text. For example, the feature quantity of text is made the ratio of kanji in the text, the ratio of katakana in the text, the ratio of hiragana in the text, the ratio of kanji to katakana in the text, the ratio of active or passive in the text, etc. Can.

  The ratio of kanji in the text is, for example, (number of kanji characters in the text) / (number of characters of the text).

  The proportion of katakana in the text is, for example, (the number of katakana characters in the text) / (the number of text characters).

  The proportion of hiragana characters in the text is, for example, (number of katakana characters in the text) / (number of characters of the fragment).

  The ratio of kanji and katakana in the text is, for example, (number of kanji and katakana characters in the text) / (number of letters of the fragment).

  The ratio of active or passive in the text is, for example, (number of appearances of active or passive in the text) / (number of appearance of verbs in the text).

<Outlier location extraction unit 4>
The outlier location extracting unit 4 receives the difficulty class of each fragment estimated by the fragment difficulty estimating unit 31 and the difficulty class of the text estimated by the overall difficulty estimating unit 32.

  The outlier location extraction unit 4 is a set of difficulty levels that is a combination of the difficulty level class of each fragment estimated by the fragment difficulty level estimation unit 31 and the difficulty level class of the text estimated by the overall difficulty level estimation unit 32. The outlier location which is a text fragment having a degree of difficulty separated from the degree of difficulty of the text estimated by the overall degree of difficulty estimation unit 32 is extracted (step S4). The extracted outlier points are output to the alternative expression presentation unit 5 and the alternative expression replacement unit 6.

  Note that the degree of difficulty, which is a combination of the difficulty level class of each fragment and the text difficulty level class, does not necessarily have to be a pair. In the following, we will explain an example of comparing in pairs, but with the difficulty class of each fragment and the text difficulty class in order to extract fragments of the text, which have difficulty different from that of the text difficulty class As long as the differences can be compared, the comparison method does not matter.

  For example, the outlier location extraction unit 4 extracts, from the difficulty level class of the text estimated by the difficulty estimation unit 3, a location that deviates by a predetermined threshold or more. That is, when the absolute value of the difference between the difficulty level class of each fragment and the estimated difficulty level class of the text is greater than or equal to a predetermined threshold (for example, 1), the outlier location extraction unit 4 The outlier location extraction unit 4 extracts each of the fragments as an outlier location.

  For example, when the difficulty level class of the estimated text is 9, "Occurrence" in the example sentences is a word that is likely to appear on difficulty level 11 or higher, as shown in FIG. The difficulty class will be 11. When the predetermined threshold is 1, 2 which is the absolute value of the difference between both difficulty levels is one or more or larger than the predetermined threshold. For this reason, the "trigger" is extracted as an outlier part.

  The outlier location extraction unit 4 may extract the outlier location using an index other than the difficulty level class such as closeness. Assuming that a fragment is a word, the closeness of a fragment that is a word is an index indicating how familiar the fragment is to be (see, for example, reference 1).

  [Reference 1] Amano Nadaaki, 1 other person, "Basic word database: Word familiarity according to meaning", Learning research company, 2008

  It is assumed that the degree of intimacy corresponding to each fragment is predetermined and stored in the storage unit 72. It is assumed that closeness corresponding to each difficulty level class is stored in the storage unit 72. FIG. 5 shows an example of familiarity corresponding to each fragment. Further, FIG. 6 shows an example of closeness corresponding to each difficulty level class. In the example of FIG. 5, the intimacy corresponding to each difficulty level class is a section. Thus, the intimacy corresponding to each overall difficulty may be a section. Moreover, in the example of FIG. 6, the intimacy corresponding to each of 12 difficulty level classes is shown by the area. In addition, on the whole difficulty level lower than the difficulty level 5, all become a very close word for adults. Therefore, the same intimacy degree may be associated with the overall difficulty level lower than the difficulty level class 5.

  The closeness degree of each fragment can be determined, for example, based on an existing database such as reference 1. The closeness can be determined, for example, in seven stages. For example, there are very few people who know that the intimacy degree is 1; if the intimacy degree is 5 or more, it indicates that 95% or more of the adults know. If the text object to be input is for an infant subject, the vocabulary acquisition age of the infant may be used to determine the intimacy.

  When the outlier location extracting unit 4 extracts outlier locations using the intimacy, the outlier location extracting unit 4 uses the intimacy corresponding to each fragment and the difficulty level class of the text estimated by the overall difficulty estimation unit 32. The corresponding closeness is read from the storage unit 7.

  For example, when the closeness degree of each fragment is lower than the closeness degree corresponding to the difficulty class estimated by the overall difficulty level estimation unit 32, the outlier point extraction unit 4 sets each fragment as an outlier point. Do.

  Thus, the outlier location extraction unit 4 extracts the outlier location based on comparison between the closeness of each fragment and the closeness corresponding to the difficulty level class of the text estimated by the overall difficulty estimation unit 32. You may

  Also, the outlier location extraction unit 4 may extract outlier locations using both of the difficulty level pair and the intimacy pair. The outlier location extraction unit 4 is, for example, the case where the absolute value of the difference between the difficulty level class of each fragment and the difficulty level class of the estimated text is a predetermined threshold (for example, 1) or more or larger. If the intimacy degree of each fragment is lower than the intimacy degree corresponding to the difficulty class estimated by the overall difficulty level estimation unit 32, each fragment is regarded as an outlier location.

  Also, as described above, the outlier location extraction unit 4 may extract the outlier location using at least one of the intimacy set and the difficulty set.

  The outlier location extracting unit 4 may extract the outlier location using at least one of the difficulty level class of each fragment and the intimacy degree of each fragment. For example, when the difficulty level class of each fragment is equal to or more than a predetermined threshold and the intimacy degree of each fragment is equal to or less than a predetermined threshold value, the outlier location extraction unit 4 sets each fragment as an outlier location. As a result, the outlier location extraction processing can be performed even on words having an extremely low frequency of appearance and words having a bias.

  FIG. 7 shows an example of the average number of words in each difficulty level class. Assuming that the difficulty level class of the input text is 3 and the number of words of a fragment which is one sentence is 15, the number of words of the fragment which is one sentence is approximately 15 and the average number of words in the difficulty level 3 of the text is approximately 11 It is greatly exceeded. For this reason, it is conceivable to use this fragment, which is one sentence, as an outlier location.

<Alternative Expression Presentation Unit 5>
Outlier points are input to the alternative expression presentation unit 5.

  The alternative expression presenting unit 5 is a fragment having the same meaning as the fragment of the outlier location extracted by the outlier location extracting unit 4, and the difficulty of the fragment of the outlier location extracted by the outlier location extracting unit 4 The user is presented with an alternative expression that is a fragment of a degree of difficulty closer to the overall degree of difficulty estimated by the overall degree of difficulty estimation unit 3 than the degree class (step S5). This presentation is performed, for example, via a display device such as a CRT or a liquid crystal display. The alternative expression presenting unit 5 may further present the outlier location to the user.

  For example, when the word “Origin” is extracted as an outlier part because the difficulty level class is too high in the outlier part extraction part 4, the alternative expression presentation part 5 has a difficulty level lower than that of the “Origin”. It is conceivable to present the word "trigger" as a candidate for replacement.

  It is assumed that the alternative expression corresponding to each fragment is stored in advance in the alternative expression storage unit 8 together with the corresponding difficulty level class. The alternative representation uses an existing database such as synonym dictionary or ALAGIN language resource / voice resource site (https://alaginrc.nict.go.jp/) where similar words and expressions are collected from a large amount of corpus. Can be created. The alternative expression presentation unit 5 reads an appropriate alternative expression by referring to the alternative expression storage unit 8 using the input outlier point as a key.

  The alternative expression presentation unit 5 may provide the user with a suggestion as to whether or not to make a correction to replace the outlier location with the presented alternative expression. When the user inputs using an input device such as a keyboard, a mouse, a touch panel or the like to accept the proposal, a correction request signal representing that is outputted from the alternative expression presentation unit 5 to the alternative expression substitution unit 6 .

  The candidate to be presented as a candidate for replacement may be one or more. In the case of multiple, the alternative expression presentation unit 5 may make a plurality of suggestions to the user whether to make corrections to replace outlier locations with any of the alternative expressions presented, or to make corrections to none of them. . When the user uses an input device such as a keyboard, a mouse, and a touch panel to input which candidate is accepted or not to correct, a correction request signal representing that is displayed on the alternative presentation presenting unit 5 Are output to the alternative expression replacing unit 6.

<Alternative Expression Replacement Unit 6>
The outlier location is input to the alternative expression replacing unit 6. Further, a correction request signal is input to the alternative representation replacing unit 6.

  The alternative expression substitution unit 6 is a fragment having the same meaning as the fragment of the outlier location extracted by the outlier location extraction unit 4, and the difficulty of the fragment of the outlier location extracted by the outlier location extraction unit 4 Of the outlier location extracted by the outlier location extracting unit 4 in the text input by the alternative expression which is a fragment of the difficulty category closer to the difficulty class estimated by the overall difficulty estimation unit 3 than the degree class The text obtained by replacing the fragment is output (step S6).

  Also, when the correction request signal is received from the alternative expression presentation unit 5, the alternative expression substitution unit 6 corrects the outlier location with the alternative expression presented by the alternative expression presentation unit 5, and corrects the corrected text Output.

  It is assumed that the alternative expression corresponding to each fragment is stored in advance in the alternative expression storage unit 8 together with the corresponding difficulty level class. The alternative representation uses an existing database such as synonym dictionary or ALAGIN language resource / voice resource site (https://alaginrc.nict.go.jp/) where similar words and expressions are collected from a large amount of corpus. Can be created. The alternative expression substitution unit 6 reads an appropriate alternative expression by referring to the alternative expression storage unit 8 using the input outlier point as a key, and performs substitution processing.

  It may be configured to output a corrected text by automatically replacing with the closest one without presenting alternative expressions. In that case, the alternative expression presentation unit 5 is unnecessary.

Second Embodiment
The outlier location extracting apparatus and method according to the second embodiment extract outlier locations from a predetermined difficulty class, not outlier locations from the difficulty class of the text estimated by the overall difficulty estimating unit 32. Apparatus and method. Hereinafter, only differences from the first embodiment will be described. The description of the same parts as in the first embodiment will be omitted.

  The predetermined difficulty level class is appropriately determined by the user.

  Unlike the outlier location extracting device according to the first embodiment, the outlier location extracting device according to the second embodiment does not include the overall difficulty estimating unit 32 for estimating the text difficulty level class, as shown in FIG. . That is, the outlier location extraction method of the second embodiment does not perform the process of step S32.

  The outlier location extracting unit 4, the alternative expression presenting unit 5 and the alternative expression replacing unit 6 of the second embodiment replace the difficulty level class of the text estimated by the overall difficulty level estimating unit 32 with a predetermined difficulty level class. Based on the same processing as the first embodiment is performed.

  That is, the outlier location extraction unit 4 of the second embodiment uses a set of difficulty levels that is a set of the difficulty level class of each fragment estimated by the fragment difficulty level estimation unit 31 and a predetermined difficulty level class. Extract outlier locations, which are fragments of the text, having difficulty levels that deviate from the predetermined difficulty level class. In other words, the outlier location extracting unit 4 of the second embodiment determines the predetermined difficulty class based on the comparison between the difficulty class of each fragment estimated by the fragment difficulty estimating unit 31 and the predetermined difficulty class. Extract outliers, which are fragments of the text, with difficulty levels away from.

  The alternative expression presentation unit 5 of the second embodiment is a fragment having the same meaning as the fragment of the outlier location extracted by the outlier location extraction unit 4, and the outlier extracted in the outlier location extraction unit 4 The user is presented with an alternative expression which is a fragment of a degree of difficulty closer to the predetermined difficulty class than the degree of difficulty of the fragment of the portion (step S5).

  The alternative expression substitution unit 6 of the second embodiment is a fragment having the same meaning as the fragment of the outlier location extracted by the outlier location extraction unit 4, and the outlier extracted in the outlier location extraction unit 4 Replace the outlier location fragment extracted by the outlier location extraction unit 4 in the text input by the alternative expression that is a fragment of the difficulty class closer to a predetermined overall difficulty level than the difficulty level class of the location fragment The output text is output (step S6).

  Also, assuming that the intimacy degree of each fragment and the intimacy degree corresponding to the predetermined difficulty level class are predetermined, the outlier location extraction unit 4 of the second embodiment determines the intimacy degree of each fragment and the predetermined difficulty level class. Outlier locations may be extracted based on the comparison with the intimacy corresponding to.

  Thus, instead of the difficulty level class estimated by the overall difficulty level estimation unit 32, a predetermined difficulty level class designated in advance by the user may be used. For example, when the user wants to set the degree of difficulty in the third grade of elementary school, a class representing the degree of difficulty in the third grade of elementary school is set as the predetermined degree of difficulty.

  For example, when it is desired to set whether the entire text is more difficult than the third grade elementary level, although the third degree primary grade is desired to be set, the overall degree of difficulty is estimated by the overall degree estimation section 32. If it is not necessary to do so, it is sufficient to simply assume a part far from the difficulty level class of the third grade of elementary school as an outlier part, so the overall difficulty level estimation unit 32 does not estimate the overall difficulty level You may

Third Embodiment
The outlier location extracting apparatus and method according to the third embodiment is an apparatus and method for estimating the difficulty level class of fragments and / or texts using the frequency of occurrence of word n-gram. Hereinafter, only differences from the first embodiment will be described. The description of the same parts as in the first embodiment will be omitted.

<Shard difficulty level estimation unit 31>
The fragment difficulty level estimation unit 31 estimates the difficulty level class of each fragment included in the input text.

  For example, if the fragment is at least one sentence (specifically, if the fragment is a paragraph, etc.), the fragment difficulty level estimation unit 31 uses the appearance frequency of each word n-gram to Estimate the difficulty class of each fragment.

  The fragment difficulty level estimation unit 31 assigns each fragment to each difficulty level class based on the appearance frequency of each word n-gram included in each fragment and the occurrence probability of each word previously obtained for each difficulty level class. The likelihood of belonging is estimated, and the highest likelihood difficulty class is taken as the difficulty class of each fragment.

  The likelihood L (i | S) that each fragment S belongs to the difficulty level class i is defined by, for example, the following equations (3 ′) and (4 ′).

Here, tf · idf (W _j ) is a weight of word n-gram W _j , f (W _j , S) is an appearance frequency of each word n-gram W _j in each fragment, and Σ _L f ( W _L , S) is the number of words n-grams contained in each of the above fragments, and D is the number of predetermined texts for learning (ie, the elements of the set D _i of texts for learning of each difficulty level number | D _i | of the sum ^{_{Σ i = 1 N | D i}} | a is), df _i is the number of learning for the text that appears in the word _{n-gram W j, P i} (W j) the degree of difficulty class The occurrence probability P _i (W _j ) of the word n-gram W _j at _i .

The occurrence probability P _i (W _j ) of the word n-gram W _j in the difficulty level class i is defined, for example, by the following equation (1). Here, i is a value (i = 1,..., N) representing the type of the difficulty level set in advance. j is a natural number of 1 or more.

Here, N is the number of difficulty classes and is a predetermined positive integer. f (w _j , D _i ) is the frequency of occurrence of fragment w _j in D _i .

P _i (w _j ) is calculated in advance as a feature amount of each word n-gram, and stored in the storage unit 71 in advance. Also, stored D required in the calculation of the formula (3 ') and Formula (4'), other parameters such as df _i in the storage unit 71. The fragment difficulty level estimation unit 31 reads these values from the storage unit 71 and calculates Equations (3 ′) and (4 ′).

The outlier location extraction device may include a precalculation unit 9 for calculating P _i (w _j ) in advance. In order to prevent the likelihood defined by the above equation (3) and equation (3 ′) from being unable to be calculated when P _i (w _j ) = 0, the pre-calculation unit 9 performs all P _i (w P _i (w _j ) may be corrected by applying the following equation (2) until _j ) is not zero.

In addition, even when P _i (w _j ) is not 0, the feature amount extraction unit 2 may use the value obtained by Expression (2). For example, even when the value obtained by equation (2) is larger than the value obtained by equation (1), the feature amount extraction unit 2 may use the value obtained by equation (2).

  In addition, a fault may arise in Formula (2) made into 1/2 of the sum of the value of both sides only in the case of 0.

For example, since P _k (w _j ) before the correction of the difficulty class d is 0, P _k (w _j ) after the correction of the difficulty class d = (0.6 + 0.6) when corrected by the equation (2). ) /2=0.3. In this case, the magnitude relationship between b and d is reversed before and after correction. That is, although P _b (w _j ) = 0.1 in the difficulty class b before correction, P _d (w _j ) = 0 in the difficulty class d, P _{b in the} difficulty class b after correction (w _j ) = 0.1 <P _d (w _j ) = 0.3 of the difficulty class d, and before and after the correction, the magnitude relationship of P _i (w _j ) is reversed between b and d. Therefore, if it has a value close to 0 and exists next to a value larger than itself, it may be corrected even if it is not 0. Specifically, in the above case, b may be corrected by the above equation (2).

Assuming that the fragment is composed of one word, that is, in the case of a word uni-gram, the occurrence probability P _i (w _j ) of the fragment w _j in the difficulty class i is “trigger” “trigger” “parent-child Etc. is an occurrence probability in each difficulty level class of each word.

  FIG. 8 shows an example of the occurrence probability in each difficulty level class of the words "trigger", "trigger" and "parent-child". Further, FIG. 9 shows an example of the occurrence probability in each difficulty level class of each of the words "Ookami", "Ookami", and "Mio". In FIGS. 8 and 9, the horizontal axis represents the difficulty level, and the vertical axis represents the probability of occurrence. In FIG. 8 and FIG. 9, there are 12 types of difficulty levels from 1 to 12. In this example, the larger the difficulty class value, the higher the difficulty. These twelve difficulty classes correspond to first graders to third graders in elementary school, respectively. That is, the difficulty level 1 corresponds to the first grader, the difficulty level 2 corresponds to the second grade, ..., and the difficulty level 12 corresponds to the third grade.

  It can be seen from FIG. 8 that the word “Origin” has a high probability of occurrence only in the high difficulty level class of 9 to 12. Also, from FIG. 9, the word “okami” has a high probability of occurrence in the low difficulty class, the word “wolf” has a high probability of occurrence in the medium difficulty class, and the word “狼” has a high difficulty class It can be seen that the occurrence probability is high.

When the fragment is composed of two words, that is, in the case of the word bigram, the occurrence probability P _i (W _j ) of the word n-gram W _j in the difficulty class i is two such as “parent-child” It is the occurrence probability in each difficulty level class of the word. As in this example, by calculating the occurrence probability when the fragment is two or more words, it is possible to calculate the occurrence probability in each difficulty level class such as a complex expression longer than a word or a wording.

  These occurrence probabilities can be used to calculate the likelihood for each difficulty level even for fragments or texts composed of longer units of n-grams.

  The fragment difficulty level estimation unit 31 may estimate the difficulty level class of each fragment by the same processing as the processing described in the first embodiment. For example, when the fragment is a word n-gram, the fragment difficulty level estimation unit 31 refers to information on the association between each fragment and the difficulty class of each fragment, which is stored in advance in the storage unit 71. You may then ask for the difficulty class of each piece.

<Overall difficulty level estimation unit 32>
The overall difficulty level estimation unit 32 estimates the difficulty level class of the input text, using the appearance frequency of each word n-gram (step S32). The degree of difficulty of the estimated text is output to the outlier location extraction unit 4.

  For example, the overall difficulty level estimation unit 32 determines the input text based on the appearance frequency of each word n-gram included in the input text and the occurrence probability of each word previously determined for each difficulty level class. The likelihood of belonging to each difficulty class is estimated, and the most likely difficulty class is taken as the difficulty class of the input text.

  The likelihood L (i | T) that the input text T belongs to the difficulty level class i is defined by, for example, the following equations (3) and (4).

Here, tf · idf (W _j ) is a weight of the word n-gram W _j , f (W _j , T) is an appearance frequency of each word n-gram W _j in T, and Σ _L f (W _L, T) is the number of words n-gram contained in the text, D is the number of predetermined training text, df _i is the number of the training text that appears in the word n-gram W _j P _i (W _j ) is an occurrence probability P _i (W _j ) of the word n-gram W _j in the difficulty class i.

P _i (W _j ) is calculated in advance as a feature amount of each word n-gram and stored in advance in the storage unit 71. Also, it stored the required D in calculation of equation (3) and (4), other parameters such as df _i in the storage unit 71. The overall difficulty level estimation unit 3 reads these values from the storage unit 71 and calculates equations (3) and (4).

The definition and pre-calculation of P _i (W _j ) are the same as described above, and thus redundant description will be omitted here.

  Note that the overall difficulty level estimation unit 32 may estimate the text difficulty level class by the same process as the process described in the first embodiment. For example, the feature extraction unit 321 in the total fragment difficulty estimation unit 32 may calculate the feature of the text, and the entire difficulty estimation unit 32 may use the calculated feature as the text difficulty class.

  The overall difficulty level estimation unit 32 estimates the likelihood for each difficulty level class based on the word n-gram, which is the result of estimating the difficulty level class of the text input using the appearance frequency of each word n-gram, The final difficulty level is estimated by the learner, using the average sentence length of the text, the ratio of kanji, the ratio of passive voice and active voice, etc. obtained by the same processing as the processing described in the first embodiment as the feature amount You may At this time, a known learner such as SVM-RANK can be used as a learner. In addition, a model for estimating the difficulty level class using these feature quantities is previously constructed from learning data and stored.

Fourth Embodiment
The outlier location extracting apparatus and method according to the fourth embodiment are different from the outlier location extracting apparatus and method according to the fourth embodiment, and the outlier location from the difficulty level class of the text estimated by the overall difficulty level estimation unit 32. And an apparatus and method for extracting outlier points from a predetermined difficulty level class. Hereinafter, only portions different from the third embodiment will be described. Descriptions of parts similar to those in the third embodiment will be omitted.

  The predetermined difficulty level class is appropriately determined by the user.

  Unlike the outlier location extracting device according to the fourth embodiment, the outlier location extracting device according to the fourth embodiment does not include the overall difficulty estimating unit 32 for estimating the difficulty level of the text as shown in FIG. . That is, the outlier location extraction method of the fourth embodiment does not perform the process of step S32.

  The outlier location extracting unit 4, the alternative expression presenting unit 5, and the alternative expression replacing unit 6 in the fourth embodiment replace the difficulty level class of the text estimated by the overall difficulty level estimating unit 32 with a predetermined difficulty level class. Based on the same processing as the third embodiment is performed.

  That is, the outlier location extraction unit 4 of the fourth embodiment uses a set of difficulty levels that is a set of the difficulty level class of each fragment estimated by the fragment difficulty level estimation unit 31 and a predetermined difficulty level class. Extract outlier locations, which are fragments of the text, having difficulty levels that deviate from the predetermined difficulty level class. In other words, the outlier location extracting unit 4 of the fourth embodiment determines the predetermined difficulty class based on the comparison between the difficulty class of each fragment estimated by the fragment difficulty estimating unit 31 and the predetermined difficulty class. Extract outliers, which are fragments of the text, with difficulty levels away from.

  The alternative expression presentation unit 5 of the fourth embodiment is a fragment having the same meaning as the fragment of the outlier location extracted by the outlier location extraction unit 4, and the outlier extracted in the outlier location extraction unit 4 The user is presented with an alternative expression which is a fragment of a degree of difficulty closer to the predetermined difficulty class than the degree of difficulty of the fragment of the portion (step S5).

  The alternative expression substitution unit 6 of the fourth embodiment is a fragment having the same meaning as the fragment of the outlier location extracted by the outlier location extraction unit 4, and the outlier extracted in the outlier location extraction unit 4 Replace the outlier location fragment extracted by the outlier location extraction unit 4 in the text input by the alternative expression that is a fragment of the difficulty class closer to the predetermined difficulty class than the difficulty class of the location fragment The output text is output (step S6).

  Further, assuming that the intimacy degree of each fragment and the intimacy degree corresponding to the predetermined difficulty level class are determined in advance, the outlier location extraction unit 4 of the fourth embodiment determines the intimacy degree of each fragment and the predetermined difficulty level class. Outlier locations may be extracted based on the comparison with the intimacy corresponding to.

  Thus, instead of the difficulty level class estimated by the overall difficulty level estimation unit 32, a predetermined difficulty level class designated in advance by the user may be used. For example, when the user wants to set the degree of difficulty in the third grade of elementary school, a class representing the degree of difficulty in the third grade of elementary school is set as the predetermined degree of difficulty.

  For example, when it is desired to set whether the entire text is more difficult than the third grade elementary level, although the third degree primary grade is desired to be set, the overall degree of difficulty is estimated by the overall degree estimation section 32. If it is not necessary to do so, it is sufficient to simply assume a part far from the difficulty level class of the third grade of elementary school as an outlier part, so the overall difficulty level estimation unit 32 does not estimate the overall difficulty level You may

[Program and Recording Medium]
When each process in the outlier location extraction device is realized by a computer, the processing content of the function that the outlier location extraction device should have is described by a program. And each process is implement | achieved on a computer by running this program by computer.

  The program describing the processing content can be recorded in a computer readable recording medium. As the computer readable recording medium, any medium such as a magnetic recording device, an optical disc, a magneto-optical recording medium, a semiconductor memory, etc. may be used.

  Further, each processing means may be configured by executing a predetermined program on a computer, or at least a part of the processing content may be realized as hardware.

[Modification]
The above outlier extraction apparatus and method may be configured to determine that the outlier location is the same as the above processing, instead of extracting the outlier location. Further, the outlier extraction apparatus and method may be configured to extract the determined outlier location after determining that it is an outlier location.

  The outlier extraction device described above includes a fragment difficulty level estimation unit 31, which is included in the input text and estimates the difficulty level of at least one of the fragments obtained by dividing the text into predetermined units, and the at least one fragment. An outlier extracting unit 4 including an outlier location extracting unit 4 that determines a fragment whose degree of difficulty exceeds a predetermined level to be an outlier location of text and / or extracts it as an outlier location of text It can be considered as an example of the device.

  The outlier location extracting unit 31 determines a fragment having a degree of difficulty which is higher than a predetermined level in at least one fragment as an outlier location of text and / or extracts it as an outlier location of text. If possible, the determination and / or extraction of outliers may be performed by processing other than the above.

  For example, when the outlier location extraction unit 31 statistically draws the distribution of the difficulty of each fragment, if the difficulty is more toward the hem than the predetermined position in the distribution, the difficulty is Determination and / or extraction of outlier locations may be performed based on statistical predetermined criteria such as setting a fragment corresponding to 外 れ as an outlier location.

  The above-mentioned difficulty class of fragments is an example of the difficulty of fragments. The difficulty levels of fragments may be discrete values or continuous values.

  In the example of FIG. 7 and FIG. 8, the number of types of the difficulty level class is 12, and the 12 difficulty levels correspond to the first grader to the third grade of elementary school, but this is an example. The number of types of difficulty classes may be N, where N is a predetermined positive integer. Also, N difficulty classes may correspond to indicators other than age and grade. For example, the difficulty level classes may correspond to special fields such as “general”, “specialty 1 (newspaper)”, “specialty 2 (patent)”, and “specialty 3 (textbook)”. This allows the outlier location extraction apparatus and method to be used when creating a manual in a particular area of expertise.

  Further, the alternative expression presentation unit 5 may not be provided in the outlier location extraction device.

  As outlier portions, not only words or paragraphs may be extracted from text, but pages may be extracted from text as outlier portions. In this case, the fragment is a page. Also, texts of different difficulty classes may be extracted from a plurality of texts as outlier locations. In this case, difficulty levels of multiple texts and difficulty levels of each text that is each fragment are estimated, and texts of difficulty levels separated from multiple text difficulty classes are extracted as outlier points . Thus, the outlier location extraction apparatus and method may be extended to various applications by changing the definitions of “entered text” and “fragment” as appropriate.

  The processes described above are not only executed chronologically according to the order of description, but may also be executed in parallel or individually depending on the processing capability of the apparatus executing the process or the necessity.

  It goes without saying that other modifications can be made as appropriate without departing from the spirit of the present invention.

1 Preprocessor 2 Feature extraction unit 31 Fragment difficulty estimation unit 32 Overall difficulty estimation unit 4 Outlier location extraction unit 5 Alternative expression presentation unit 6 Alternative expression replacement unit 71 Storage unit 72 Storage unit 8 Alternative expression storage unit

Claims (10)

A fragment difficulty level estimation unit for estimating the difficulty level class of each fragment obtained by dividing the text into predetermined units, which is included in the input text;
An overall difficulty estimation unit that estimates the difficulty level class of the above text,
A fragment of the text having a degree of difficulty separate from the estimated difficulty level of the text based on a comparison of the estimated difficulty level of each fragment and the estimated difficulty level of the text An outlier location extraction unit for extracting a certain outlier location;
Only including,
The fragment difficulty level estimation unit determines each difficulty level of each fragment based on the appearance frequency of each word n-gram included in each fragment and the occurrence probability of each word previously obtained for each difficulty level class. The likelihood belonging to the class is estimated, and the highest likelihood difficulty class is taken as the difficulty class of each fragment estimated above,
The overall difficulty level estimation unit determines whether the text is in each difficulty level class based on the appearance frequency of each word n-gram included in the text and the occurrence probability of each word previously determined for each difficulty level class. Estimate the likelihood of belonging, and let the difficulty class with the highest likelihood be the difficulty class of the above estimated text,
Outlier location extractor. A fragment difficulty level estimation unit for estimating the difficulty level class of each fragment obtained by dividing the text into predetermined units, which is included in the input text;
An outlier portion which is a fragment of the text having a degree of difficulty separated from the predetermined degree of difficulty is extracted based on a comparison between the estimated degree of difficulty of each fragment and a predetermined degree of difficulty. Outlier location extraction unit,
Only including,
The fragment difficulty level estimation unit determines each difficulty level of each fragment based on the appearance frequency of each word n-gram included in each fragment and the occurrence probability of each word previously obtained for each difficulty level class. The likelihood belonging to the class is estimated, and the highest likelihood difficulty class is taken as the difficulty class of each fragment estimated above,
Outlier location extractor. In the outlier location extraction device according to claim 1 or 2 ,
Let S be the fragment, f (W _j , S) be the appearance frequency of each word n-gram W _j in each fragment, and _{L L} f (W _L , S) be the word n-gram contained in each fragment Let D be the number of given training texts, d f _{i be} the number of the above training texts in which word n-gram W _j appears, and P _i (W _j ) be the word n- in the difficulty class i As the occurrence probability P _i (W _j ) of gram W _j, the likelihood L (i | S) that each of the above fragments belongs to the difficulty level class i is defined by the following equation

Let T be the above text, f (W _j , T) be the appearance frequency of each word n-gram W _j in T, and _{L L} f (W _L , T) be the number of word n-grams contained in the text Let D be a predetermined number of learning texts, d f _{i be} the number of the learning texts in which word n-gram W _j appears, and P _i (W _j ) be a word n-gram W in difficulty level i _As the occurrence probability P _i (W _j ) of _j , the likelihood L (i | T) that the above text belongs to the difficulty level class i is defined by the following equation:

Outlier location extractor. In the outlier location extraction device according to any one of claims 1 to 3 ,
Assuming that the above fragment is a word,
The intimacy degree of a word is used as an index indicating how familiar the word is felt, and corresponds to the intimacy degree of each fragment and the difficulty level of the estimated text or the predetermined difficulty level. Assuming that closeness is predetermined,
The outlier location extraction unit is configured to compare the outlier location based on a comparison between the intimacy degree of each fragment and the intimacy degree corresponding to the estimated difficulty level of the text and / or the predetermined difficulty level class. Extract,
Outlier location extractor. In the outlier location extraction device according to any one of claims 1 to 4 ,
The fragment is at least one sentence, and the feature quantity of the fragment is the average number of words constituting the fragment, or
The above fragment is at least one sentence, and the feature quantity of the fragment is the number of clauses constituting the fragment, or
The fragment is at least one sentence, and the feature quantity of the fragment is the ratio of kanji in the fragment,
The fragment is at least one sentence, and the feature quantity of the fragment is the proportion of katakana in the fragment, or
The fragment is at least one sentence, and the feature quantity of the fragment is the percentage of hiragana in the fragment,
The above fragment is at least one sentence, and the feature quantity of the fragment is the ratio of kanji and katakana in the fragment,
The fragment is at least one sentence, and the feature quantity of the fragment is the rate of activity or passiveness of the fragment,
The above fragment is at least one sentence, and the feature quantity of the fragment is the depth of the syntax tree in the fragment, or
An outlier location extractor that is any of the following. In the outlier location extraction device according to any one of claims 1 to 5 ,
It is a fragment having the same meaning as the extracted fragment of the outlier part, which is the difficulty class of the above-mentioned estimated text or the above predetermined difficulty class than the difficulty of the fragment of the extracted outlier part An alternative presentation presenting unit that presents to the user an alternative presentation that is a fragment of a degree of difficulty close to
An outlier location extractor further comprising In the outlier location extraction device according to any one of claims 1 to 5 ,
It is a fragment having the same meaning as the extracted fragment of the outlier part, which is the difficulty class of the above-mentioned estimated text or the above predetermined difficulty class than the difficulty of the fragment of the extracted outlier part An alternative expression substitution presentation unit which outputs a text obtained by replacing the extracted fragment of the outlier location in the text by an alternative expression that is a fragment having a degree of difficulty near,
An outlier location extractor further comprising   A fragment difficulty level estimation step of estimating a difficulty level class of each fragment obtained by dividing the text into predetermined units, which is included in the input text;
  An overall difficulty estimation step of estimating the difficulty level of the above text by the overall difficulty estimation unit;
  The outlier location extraction unit determines the degree of difficulty away from the estimated difficulty class of the text based on the comparison of the estimated difficulty class of each fragment with the estimated difficulty class of the text. An outlier location extracting step of extracting an outlier location which is a fragment of the text,
  Including
  In the fragment difficulty level estimation step, each fragment has each difficulty level based on the appearance frequency of each word n-gram included in each fragment and the occurrence probability of each word previously determined for each difficulty level class The likelihood belonging to the class is estimated, and the highest likelihood difficulty class is taken as the difficulty class of each fragment estimated above,
  In the overall difficulty level estimation step, the text is divided into each difficulty level class based on the appearance frequency of each word n-gram included in the text and the occurrence probability of each word previously determined for each difficulty level class. Estimate the likelihood of belonging, and let the difficulty class with the highest likelihood be the difficulty class of the above estimated text,
  Outlier location extraction method.   A fragment difficulty level estimation step of estimating a difficulty level class of each fragment obtained by dividing the text into predetermined units, which is included in the input text;
  The text fragment of the above, wherein the outlier location extraction unit has a difficulty level away from the predetermined difficulty level class based on the comparison between the estimated difficulty level class of each fragment and the predetermined difficulty level class An outlier location extracting step for extracting a certain outlier location;
  Including
  In the fragment difficulty level estimation step, each fragment has each difficulty level based on the appearance frequency of each word n-gram included in each fragment and the occurrence probability of each word previously determined for each difficulty level class The likelihood belonging to the class is estimated, and the highest likelihood difficulty class is taken as the difficulty class of each fragment estimated above,
  Outlier location extraction method. A program for causing a computer to function as each part of the outlier location extraction device according to any one of claims 1 to 7 .

Images