JP5706379B2 - Feature detection apparatus, feature detection method and program thereof - Google Patents

Description

  The present invention relates to a technique for detecting features of a vocabulary learning curve.

  Although human language development can provide important scientific knowledge and suggestions for thinking about "what is human beings", there are currently many unresolved issues regarding language development. For this reason, there is almost no progress in measurement technology related to language development and commercial service development. In particular, there has been little progress in science and technology related to vocabulary acquisition, among basic abilities such as speech recognition, vocabulary acquisition, and grammar manipulation. However, considering the need for education that moderately supports healthy development and the early detection and support of developmental disabilities, including language development delays, technical development related to vocabulary acquisition is considered to be important.

  One of the most characteristic and important aspects of preschool language development is the vocabulary learning curve (the age at which an infant begins to speak a new word, Graph showing the relationship with the cumulative number of words that have come to be. The characteristics of this vocabulary learning curve cannot be obtained without the phenomenon called vocabulary explosion (or vocabulary spurt). Vocabulary explosion is a phenomenon that development psychologists have been focusing on since the middle of the twentieth century and refers to a rapid change in vocabulary learning speed that occurs in the second half of the year. Basically, infants begin to utter their first words around their first birthday, but for a while after that they learn words at a very moderate rate. However, after the age of one and a half years, the word suddenly starts to utter, so the dramatic change is called "explosion" or "spurt". The vocabulary explosion is so dramatic that many parents consciously notice it, so it is well known not only in the field of psychology but also in the childcare industry. For this reason, it is considered that vocabulary explosions must be modeled in order to quantify the state of child vocabulary development (learning).

  Conventionally, in the field of developmental psychology, the phenomenon of vocabulary explosion has been confirmed in multiple languages using large-scale group data using a vocabulary checklist (questionnaire survey based on parents' answers). It has been found that when the average value of the group data is plotted for each age, it becomes a quadratic curve showing a gradual increase, and its inflection point appears around 18-20 months. From these collective data, we consider that the vocabulary explosion is a common phenomenon seen in many children.

  As a method of examining the characteristics of the vocabulary learning curve, the vocabulary explosion time is determined from the viewpoint of when the vocabulary explosion occurs for each individual and how to detect and estimate the vocabulary explosion time (the time when the vocabulary explosion starts). The following six methods have been proposed.

  (1) A visual method for determining a vocabulary explosion time by drawing a graph without any particular calculations.

  (2) A method of approximating a vocabulary learning curve with a quadratic curve showing a gradual rise and setting an inflection point appearing around 18-20 months as a vocabulary explosion time.

  (3) The 50-word achievement standard method that defines the time when 50 words are learned as the vocabulary explosion time.

  (4) A specific period achievement standard method in which a vocabulary explosion period is defined as a period that satisfies an achievement standard (for example, 30 words or more) in a specific period (for example, three weeks)

  (5) A logistic regression approximation method in which the velocity component of the vocabulary acquisition data along the time axis is approximated to a logistic regression equation, and the inflection point is the vocabulary explosion time (see Non-Patent Document 1).

  (6) A method of approximating a vocabulary acquisition line by two straight lines before and after the vocabulary explosion, minimizing the sum of the errors, and setting the intersection as the vocabulary explosion time (see Non-Patent Document 2).

  Until now, the vocabulary learning curve was determined by determining the vocabulary explosion time using these methods, obtaining the vocabulary learning speed before and after that time, or using the approximate curve used to determine the vocabulary explosion time. I knew the characteristics.

Ganger, J., Brent, M. R., "Reexamining the vocabulary spurt", Developmental Psychology, 2004, Vol.40, No.4, p.621-632. Yasuhiro Minami, Tetsuo Kobayashi, Hiroaki Sugiyama, “Estimation of the Start of Vocabulary Explosion by Line Approximation”, IEICE Transactions, March 2012

  In the methods (1) to (4), the data of a large number of infants are averaged to determine the timing of one vocabulary explosion, which is used as a feature of the vocabulary learning curve. In such a process, since the data is averaged, the characteristic of the vocabulary learning curve unique to each infant cannot be obtained.

  In the methods (5) and (6), by using the data for each individual, it is possible to obtain more specific features unique to each infant. However, since these methods also assume that the timing of vocabulary explosion is one point or that there is no vocabulary explosion, it is not possible to obtain more detailed characteristics of the vocabulary learning curve. The reason why the detailed features cannot be handled by the methods (5) and (6) will be described using the actual vocabulary learning curve of each infant. FIG. 1 shows the vocabulary learning curve of the infant 1 and FIG. In these figures, the horizontal axis indicates the infant's age, and the vertical axis indicates the cumulative number of learned words. A complex vocabulary learning curve in which a large discontinuity appears such as an S-shaped curve as shown by a broken line in FIG. 1 and a large discontinuity as shown by a broken line in FIG. It is difficult to model with one vocabulary explosion by this method.

  Therefore, in the present invention, the characteristics of the vocabulary learning curve are obtained from a different viewpoint from the conventional methods. An object of the present invention is to provide a feature detection apparatus, a feature detection method, and a program for detecting a feature of a vocabulary learning curve different from a vocabulary explosion without assuming a vocabulary explosion.

  In order to solve the above-described problem, according to the first aspect of the present invention, the feature detection apparatus is configured so that the infant speaks a new word and the infant speaks before the age. Based on the vocabulary learning curve that shows the relationship with the cumulative number of words that have become, a first plateau detection unit that detects a portion where the age interval when a new word is spoken is greater than a threshold, and a vocabulary learning curve And a second plateau detection unit that detects a portion where the vocabulary learning speed of the smoothed vocabulary learning curve is minimized.

  In order to solve the above-described problem, according to a second aspect of the present invention, a feature detection method in a feature detection apparatus including a first plateau detection unit and a second plateau detection unit is performed by an infant using a first plateau detection unit. Utters new words based on a vocabulary learning curve that shows the relationship between the age at which a child began speaking a new word and the cumulative number of words spoken by an infant by age A first plateau detection step for detecting a portion where the age interval is larger than a threshold value, and a portion where the vocabulary learning speed of the smoothed vocabulary learning curve is minimized by smoothing the vocabulary learning curve by the second plateau detection unit Detecting a second plateau.

  According to the present invention, it is possible to detect a plateau that is a feature of a vocabulary learning curve different from a vocabulary explosion.

The figure showing the vocabulary learning curve of the infant 1. The figure showing the vocabulary learning curve of the infant 2. The figure which shows the position of the plateau of the vocabulary learning curve of the infant 1. The figure which shows the position of the plateau of the vocabulary learning curve of the infant 2. The functional block diagram of the feature detection apparatus which concerns on 1st embodiment. The figure which shows the processing flow of the feature detection apparatus which concerns on 1st embodiment. The figure which shows the processing flow of the 1st plateau detection part 110. FIG. The figure which shows the vocabulary learning curve of the infant 1 which made the cumulative number i of the word a horizontal axis. The figure which shows the series of the days required in order to learn 1 word of the vocabulary learning curve obtained when smoothing the vocabulary learning curve of the infant 1. The figure which shows the part where the number of days required in order to learn 1 word of the infant 1 becomes the maximum. The figure which shows simultaneously the vocabulary learning curve of the infant 1, and the part where the number of days required to learn one word becomes the maximum. The figure which shows the vocabulary learning curve of the infant 2 with the horizontal axis | shaft as the cumulative number i of a word. The figure which shows the series of the days required in order to learn 1 word of the vocabulary learning curve obtained when smoothing the vocabulary learning curve of the infant 2. The figure which shows the part in which the number of days required to learn one word of the infant 2 becomes the maximum. The figure which shows simultaneously the vocabulary learning curve of the infant 2, and the part where the number of days required to learn one word becomes the maximum. The figure which shows the processing flow of a merge part. The figure which shows the position of the plateau which merged FIG. 3 and FIG. The figure which shows the position of the plateau which merged FIG. 4 and FIG.

  Hereinafter, embodiments of the present invention will be described. In the drawings used for the following description, constituent parts having the same function and steps for performing the same process are denoted by the same reference numerals, and redundant description is omitted.

<Summary of invention>
First, the inventors discovered that this is not a simple phenomenon in which the vocabulary rapidly increases at a time when there is a vocabulary explosion, but a phenomenon in which several vocabulary increases overlap (references). 1).
[Reference 1] Y. Minami, H. Sugiyama, T. Kobayashi, "Multiple Vocabulary Spurts in Japanese Children", 12th International Congress for the Study of Child Language (IASCL2011), 2011

  Furthermore, if we look closely at the vocabulary learning curve, we found that these increases can be expressed in a simple curve (straight line or quadratic curve) by a model that inserts multiple intervals where new words are not spoken for more than a few days. The section (part) in which this new word is not spoken for more than a few days is named plateau (a technical term in learning psychology that means “plain”). In fact, it can be seen from FIG. 1 and FIG. 2 that there are plateaus (some of which are shown in the part surrounded by a solid line). Since this modeling assumes that the original vocabulary learning curve is a very simple curve (straight line or quadratic curve), these plateaus are key features for understanding the phenomenon of vocabulary explosion. This is a characteristic of the vocabulary learning curve.

As a simple method of finding the plateau position, a method of finding a period in the vocabulary learning curve that is longer than a threshold value p (for example, p = 6) and in which no word is learned can be considered. Here, when y _i is the age at which the i-th word is learned, when y _i −y _i−1 > p, y _i−1 to y _i −1 are plateaus, and y _i−1 is a plateau. And y _i −1 is the end position of the plateau. In other words, the period from the cumulative number i-1 to the previous day from the cumulative number i is called a plateau. By such a method, the plateau can be easily detected as shown in FIGS. 3 and 4 are displayed by replacing the y-axis and the x-axis in FIGS. 1 and 2, respectively. That is, the y-axis is age, and the x-axis is the cumulative number of words. In the figure, the position indicated by □ corresponds to the end position of the plateau. More specifically, the day before the position indicated by □ is the end position of the plateau.

  Although the human eye can confirm that the vocabulary learning curve is stepped like the part surrounded by the solid line in FIG. 4, the interval is smaller than the threshold value p, so it is detected as a plateau. There is a section (part) that is not performed. If such a section (part) is called a plateau candidate, it can be seen that a considerable number of plateau candidates are observed in the vocabulary learning curve.

  This plateau candidate is out of the definition of the plateau described above. However, in the sense that the vocabulary learning curve changes sharply and the vocabulary learning speed decreases rapidly, this is considered to be a point that bears the characteristics of the vocabulary learning curve. Therefore, the definition of the plateau is extended to a section (part) where the vocabulary learning speed is minimized.

  That is, the plateau is defined as (i) a section (part) longer than a predetermined period (threshold) and not speaking a new word, or (ii) a section (part) where the vocabulary learning speed is minimized. When the vocabulary learning speed is minimized, (a) the vertical axis represents the age, and the horizontal axis represents the cumulative number of words. The lexical learning curve has a steep slope, and (b) the vertical axis represents the cumulative number of words. When the horizontal axis is the age, the slope of the vocabulary learning curve becomes gentle.

  In the embodiment, an apparatus for detecting this plateau will be described.

<First embodiment>
FIG. 5 is a functional block diagram of the feature detection apparatus 100 according to the first embodiment, and FIG. 6 shows a processing flow thereof.

  The feature detection apparatus 100 includes a first plateau detection unit 110, a second plateau detection unit 120, and a merge unit 130. The second plateau detection unit 120 includes a smoothing unit 121 and a detection unit 123.

Feature detector 100 includes infants and Nichiyowai y _i which is adapted to utter the new word, a data set consisting of a cumulative number i of words infants up to that Nichiyowai y _i began to speech {(1 , Y ₁ ), (2, y ₂ ),..., (I, y _i ),..., (I, y _I )} and receive a set of plateau positions {r ₁ , r _{2 ,.} ..., r _K } is obtained and output. K is the number of plateaus included in the vocabulary acquisition curve. The data sets {(1, y ₁ ), (2, y ₂ ),..., (I, y _i ),..., (I, y _I )} are as shown in FIGS. Since it represents a vocabulary learning curve for each infant, it is also simply called a vocabulary learning curve. Also, i may be said to be the order of words newly learned by the infant, and y _i may be said to be the age at which the i-th word was spoken. The outline of this embodiment will be described below.

  In the first plateau detection unit 110, (i) a section (part) that is longer than a predetermined period (threshold) and does not speak a new word is detected.

The second plateau detection unit 120 detects (ii) a section (part) where the vocabulary learning speed (the differential coefficient of the vocabulary learning curve) is minimized. Here, since the observed value includes randomness, if the minimum value of the differential coefficient is obtained directly from the vocabulary learning curve, many minimum points are generated. Therefore, in this embodiment, the Kalman filter called a trend model is used to smooth the vocabulary learning curve and calculate the differential coefficient. At this time, the vertical axis and the horizontal axis in FIGS. 1 and 2 are interchanged. That is, as shown in FIGS. 3 and 4, the horizontal axis represents the cumulative number, and the vertical axis represents the age. This processing is introduced in order to meet the demand that the time intervals of the Kalman filter need to be arranged evenly. However, in order to find the plateau, instead of differentiating the data series of the cumulative number i by the age y _i and detecting the local minimum, it is necessary to differentiate the data series of the age y _i by i and detect the local maximum There is.

  In the merging unit 130, the plateaus detected by the first plateau detection unit 110 and the second plateau detection unit 120 are merged and output.

  In addition, it is known from the observation of the vocabulary learning curve by the inventors so far that the interval between the plateaus tends to become shorter as the age increases. With the above-described configuration, the first plateau detection unit 110 can detect a plateau at the initial stage of vocabulary learning, and the second plateau detection unit 120 can detect a plateau after the middle period of vocabulary learning. Details will be described below.

(About collecting data sets)
In order to estimate the plateau in the infant's vocabulary learning curve, the first question is what data to refer to. If it is possible to record all of the infant's utterances with electronic media such as a digital video recorder, it can be said that the most accurate method is to analyze it, but the cost of data acquisition is enormous and the infant's ambiguous speech data Since there is no engineering technology that automatically recognizes and analyzes at the word level, it is very difficult to realize. On the other hand, there is also a method in which a questionnaire is answered every predetermined period (for example, once every three months) to grasp the change in the number of words newly spoken by the infant. In this case, if the predetermined period is long, it is difficult to grasp the exact timing of the vocabulary explosion. In addition, if the predetermined period is short, the burden on the respondent of the questionnaire (the parent of the infant) increases. Therefore, in practice, it is desirable to have a method capable of acquiring data at fine time points while reducing the burden on the parent who records the data.

Therefore, in this embodiment, data acquisition using the web diary method is applied. In this method, when an infant learns a new word (speaks), a specific site on the web is accessed via a network from a mobile phone or a personal computer, and the word that the infant remembers is recorded along with the diary of the day. (See References 2 and 3).
[Reference 2] Tetsuo Kobayashi, Masaaki Nagata, “Infant Language Development Research Using the Web: A Trial of Large Scale Longitudinal Data Collection”, Proc. 15th Annual Conference of the Language Processing Society, 2009, p. 534-537
[Reference 3] Tetsuo Kobayashi, Masaaki Nagata, “Reliability Verification of Infant Vocabulary Development Data Collected on the Web”, Proc. Of the 16th Annual Conference of the Language Processing Society, 2010, p. 403-406
The effectiveness of this method is very good in that it has been scientifically verified.

  In addition, the advantage of data acquisition by this method is that it is relatively easy for parents to record, but the date of recording (the date when the infant learned a new word) and the date of birth of the infant From the difference, it is possible to calculate the age at which the infant learned a new word.

  The calculated acquisition ages of each word are arranged in ascending order, and an integer series of 1, 2,... Is assigned from the smallest to calculate the cumulative number i of words. Here, i represents the cumulative number of words and the order in which the corresponding words are learned. Thereby, the data set which consists of a set of age and the cumulative number of words is produced | generated.

In this embodiment, direct, word data set consisting cumulative number i and day-old _{y i {(1, y 1} ), (2, y 2), ..., (i, y i), ..., (I, Suppose y _I )} is given as input.

<First Plateau Detection Unit 110>
The first plateau detector 110, the data set _{{(1, y 1),} (2, y 2), ..., (i, y i), ..., (I, y I)} receives the word infant new A portion where the interval (y _i -y _i-1 ) at which the new word starts to be spoken is greater than the threshold p based on the age y _{i at} which the new word starts to be spoken (s110). sets of sets plateau position of cumulative number i of words corresponding to the end position of the part _{{r 1,1, r 1,2, ...} , r 1, v, ..., r 1, V} combining unit 130 as Output to. When the age interval (y _i -y _i-1 ) is larger than the threshold value p, the cumulative number of plateau start positions is i-1, and the cumulative number corresponding to the end position is i. The cumulative number i is not the cumulative number of the plateau end position, but the cumulative number corresponding to the plateau end position. And y _i -1).

  FIG. 7 shows an example of the processing flow of the first plateau detection unit 110.

First, initialization is performed (s111). Next, it is determined whether or not the age interval (y _i −y _i−1 ) at which a new word is spoken is larger than a threshold value p (s112). If it is larger, the cumulative number i of words is substituted into the plateau positions r _{1 and v} (s113), and v is incremented (s114).

  The processing of s112 to s114 is performed for all the cumulative numbers i (s115, s116), and then the number of plateaus v-1 detected by the first plateau detection unit 110 is substituted for V (s117).

  For example, the cumulative number i of words corresponding to the position of the plateau detected by the first plateau detection unit 110 is the cumulative number of words corresponding to the positions indicated by the squares in FIGS.

<Second Plateau Detection Unit 120>
The second plateau detection unit 120 includes a smoothing unit 121 and a detection unit 123. The second plateau detection unit 120 smoothes the vocabulary learning curve, detects a portion where the vocabulary learning speed of the smoothed vocabulary learning curve is minimized, and determines the cumulative number i of words corresponding to that portion as the plateau position r. _{2 and v} are output to the merging unit 130. Hereinafter, processing contents of the smoothing unit 121 and the detection unit 123 will be described.

(Smoothing part 121)
The smoothing unit 121 receives the data set {(1, y ₁ ), (2, y ₂ ), ..., (i, y _i ), ..., (I, y _I )}, and removes the influence of noise. Is used to smooth the vocabulary learning curve represented by the data set {(1, y ₁ ), (2, y ₂ ), ..., (i, y _i ), ..., (I, y _I )} ( Smoothing) (s121).

  For example, a Kalman smoother represented by the following equation of state is used as a filter.

_{Here, v 1, i, v 2} , i and _{w i} are both a variable generated from the Gaussian distribution, both mean 0, variance, respectively 0.05, 0.05 and 0.1 is there. By performing the Kalman smoother using the state equation, the day data series (vocabulary learning curve) age _{{y 1, y 2, ...} , y i, ..., y I} by smoothing, smooth Kana Date age data series (vocabulary learning _{curve) {x 1, x 2,} ..., x i, ..., x I} obtained.

If using the Kalman smoother, the number of days required to learn one word, is assigned to δx _i. This value δx _i corresponds to the differential coefficient of the vocabulary learning curve. Therefore, the resultant vocabulary learning curve _{{x 1, x 2, ...} , x i, ..., x I} without differentiating the directly one series of days required to learn the words {.delta.x _1, δx ₂ ,..., δx _i ,..., δx _I } can be obtained.

However, without using the value .delta.x _i, smooth vocabulary learning curve _{{x 1, x 2, ...} , x i, ..., x I} differentiates, it may be configured to determine the differential coefficient. As described above, the differential coefficient corresponds to the number of days δx _i needed to learn one word. Further, a smooth vocabulary learning curve {x ₁ , x ₂ ,..., X _i ,..., X _I } is used to calculate the difference in days (δx _i = x _i −x _i−1 ), that is, to speak one word. You may ask for the required number of days.

The smoothing unit 121 includes a data set {(1, δx ₁ ), (2, δx ₂ ) including the number of days δx _i necessary to learn one word obtained by any method and the cumulative number i of words. ,..., (I, δx _i ),..., (I, δx _I )} are output to the detection unit 123.

(Detector 123)
The detection unit 123 receives the data set {(1, δx ₁ ), (2, δx ₂ ),..., (I, δx _i ),..., (I, δx _I )} to learn one word. From the series of necessary days {δx ₁ , δx ₂ ,..., Δx _i ,..., Δx _I }, a portion where the number of days δx _i necessary to learn one word is maximized is detected (s123). set of plateau position the set of cumulative number i of words corresponding to _{{r 2,1, r 2,2, ...} , r 2, w, ..., r 2, W} and outputs the combining unit 130 as. The portion of the number of days .delta.x _i needed to learn a word becomes the maximum is a portion inclined vocabulary learning curve becomes steeper (plateau) is a portion vocabulary learning rate is minimized.

  8 and 12 are the vocabulary learning curves for infant 1 and infant 2, respectively, with the horizontal axis representing the cumulative number of words i, and FIGS. 9 and 13 are for smoothing the vocabulary learning curves for infant 1 and infant 2, respectively. FIG. 10 and FIG. 14 show the maximum number of days required to learn one word for infant 1 and infant 2, respectively. 11 and 15 simultaneously show the vocabulary learning curve of infant 1 and infant 2 and the portion where the number of days required to learn one word is maximized.

  It can be seen from FIG. 15 that the stepped portion of FIG. 4 has been extracted.

Since i is a discrete value, the maximum value of δx _i can be obtained by an easy calculation. For example, when the number of days δx _i necessary to learn one word is larger than both the number of days δx _i−1 and δx _{i + 1} necessary to learn one adjacent word, that one word is learned. a method of the maximum value dates .delta.x _i necessary, the number of days of sequences needed to learn a word _{{δx 1, δx 2, ...} , δx i, ..., δx I} determined an approximate expression for, A method for obtaining a local maximum value using the Newton method or the like for the approximate expression is conceivable.

<Consolidation unit 130>
Merging unit 130, a set of plateaus positions _{{r 1,1, r 1,2, ...} , r 1, v, ..., r 1, V} and _{{r 2,1, r 2,2, ...} , r _2, w, ..., receive and _{r 2, W},} were combined to remove the overlapping portion (s130), the set _{{r 1} position merged _{plateau, r 2, ..., r k} , ..., r K } As an output value of the feature detection apparatus 100.

  FIG. 16 shows an example of the processing flow of the merging unit 130.

First, initialization is performed (s130a). Next, the sizes of r1 _{, v} and r2 _{, w} are compared (s130b, s130h).

When r _{1 and v} are smaller (s 130 b), r _{1 and v} are substituted into r _k (s 130 c), k and v are incremented (s 130 d), and v is detected by the first plateau detection unit 110. It is determined whether or not the number of plateaus is greater than V (s130e). If it is V or less, the process returns to s130b. If larger than V, the following process (1) is performed (s130f), k-1 is substituted for K (s130g), and the merge process is terminated. In process (1), all the remaining elements of r _{2 and w} are added to the set of plateau positions. _{Specifically,} by substituting _{r 2, w} to _{r k (s130f-1),} k is incremented and w (s130f-2), w is the number of plateaus that are detected by the second plateau detector 120 W Until it becomes larger, s130f-1 and s130f-2 are repeated (s130f-3).

When r _{2 and w} are smaller (s 130 b and s 130 h), r _{2 and w} are substituted into r _k (s 130 i), k and w are incremented (s 130 j), and w is detected by the second plateau detection unit 120. It is determined whether or not the number of plateaus made is greater than W (s130k). If it is less than or equal to W, the process returns to s130b. If larger than W, the following process (2) is performed (s130l), k-1 is substituted for K (s130g), and the merge process is terminated. In process (2), all the remaining elements of r _{1 and v} are added to the set of plateau positions. _{Specifically,} by substituting _{r 1, v} to _{r k (s130l-1),} k is incremented and v (s130l-2), v a is the number of plateaus that are detected by the first plateau detector 110 V Until it becomes larger, s130l-1 and s130l-2 are repeated (s130l-3).

If r _{1, w} and _{r 2, w} and is the same (s130b, s130h), one of the values (e.g. _{r 2, w)} is substituted into _{r k (s130m),} increments k, v and w (S130n), whether v is the number of plateaus detected by the first plateau detection unit 110 larger than V, and whether w is the number of plateaus detected by the second plateau detection unit 120 larger than W. (S130o, s130p, s130r). Note that processing equivalent to deleting duplicate values by incrementing a plateau index (v in this example) that has not been subjected to substitution processing is performed. If v> V and w> W, k-1 is substituted for K (s130g), and the merging process is terminated. If v ≦ V and w ≦ W, the process returns to s130b. When v> V and w ≦ W, processing (1) is performed (s130q), k-1 is substituted for K (s130g), and the merging processing is terminated. When v ≦ V and w> W, processing (2) is performed (s130s), k-1 is substituted for K (s130g), and the merging processing is terminated.

  FIGS. 17 and 18 show the positions of the plateaus obtained by merging FIGS. 3 and 11 and FIGS. 4 and 15 respectively.

  The plateau cannot be detected at the initial stage of word learning where the plateau is frequently inserted only by the processing of the second plateau detection unit 120. Therefore, the final plateau is obtained by merging the plateau obtained by the processing of the first plateau detection unit 110 in the merging unit 130.

<Effect>
With such a configuration, it is possible to detect a plateau that is a feature of a vocabulary learning curve different from the vocabulary explosion. The plateau can be automatically presented via an output unit (display, printer, etc.) not shown. It is possible to extract the vocabulary acquisition characteristics of each individual from the infant-specific plateau included in the vocabulary learning curve. This can ultimately make tailored education tailored to development more effective and can be a commercially valuable indicator.

  In addition, when the plateau occurs, there is a high possibility that an important phenomenon has occurred in the vocabulary learning process, and finding this time is important for understanding the vocabulary learning process scientifically.

<Modification>
In this embodiment, data is directly input / output between each unit, but data may be input / output via a storage unit (not shown).

  As the form of the Kalman smoother, the above-described equation of state is used. However, any equation can be used as the equation of state as long as it can calculate the trend of the data series of days of age. However, depending on the state equation, the state equation may not include a variable for the number of days taken to utter one word. In this case, the number of days required to speak one word can be calculated by obtaining the differential coefficient of the vocabulary learning curve smoothed by the filter. Here, the result of the Kaman smoother by the state equation is used as the filter result, but the result of the Kalman filter using the same state equation may be used.

  Further, although the state equation is used as the shape of the filter, other smoothing methods such as a low-pass filter may be used. Also in this case, the number of days required to speak one word can be calculated by obtaining the differential coefficient of the vocabulary learning curve smoothed by the filter.

  In this embodiment, the data series of ages arranged in the order of the cumulative number is smoothed, but the data series of cumulative numbers arranged in the order of ages may be smoothed. That is, smoothing may be performed with the vertical axis as the cumulative number of words and the horizontal axis as the age. However, in that case, in order to perform the filtering using the Kalman smoother or the like, the time intervals need to be evenly arranged, so it is necessary to supplement the cumulative number at each age. Since the number of data and processing increase due to complementation, the first embodiment is more advantageous. In the first embodiment, since the supplemented value is not used, the detection accuracy is considered to be equal to or higher than that in the case where the supplemented value is used.

  In the present embodiment, the first plateau detection unit 110 uses the cumulative number i corresponding to the end position as the plateau position, but may be configured to use the cumulative number i−1 of the start position.

  Further, the merging unit 130 may perform the merging process using another algorithm. For example, the merge processing may be performed using a function Union such as Lisp (a kind of programming language), matlab (registered trademark), R language, or the like. Furthermore, the merging unit 130 may not be provided. In that case, the plateau detected by the first plateau detector 110 and the plateau detected by the second plateau detector 120 may be output as they are, or a label indicating which plateau is detected by which detector is used. You may add and output.

  The present invention is not limited to the above-described embodiments and modifications. For example, the various processes described above are not only executed in time series according to the description, but may also be executed in parallel or individually as required by the processing capability of the apparatus that executes the processes. In addition, it can change suitably in the range which does not deviate from the meaning of this invention.

<Program and recording medium>
The feature detection apparatus described above can also be functioned by a computer. In this case, each process of a program for causing a computer to function as a target device (a device having the functional configuration shown in the drawings in various embodiments) or a process procedure (shown in each embodiment) is processed by the computer. A program to be executed by the computer may be downloaded from a recording medium such as a CD-ROM, a magnetic disk, or a semiconductor storage device or via a communication line into the computer, and the program may be executed.

100 feature detection device 110 first plateau detection unit 120 second plateau detection unit 121 smoothing unit 123 detection unit 130 merging unit

Claims (5)

Speak a new word based on a vocabulary learning curve that shows the relationship between the age at which an infant began speaking a new word and the cumulative number of words spoken by the infant by that age A first plateau detection unit that detects a portion where the interval of the age of becoming larger than a threshold value,
A second plateau detection unit that smoothes the vocabulary learning curve and detects a portion where the vocabulary learning speed of the smoothed vocabulary learning curve is minimized.
Feature detection device. The feature detection device according to claim 1,
The second plateau detector is
A smoothing unit that smoothes the vocabulary learning curve using a Kalman smoother;
A detection unit that detects a portion where the vocabulary learning speed is minimized from a sequence of days required to learn one word obtained when performing smoothing,
Feature detection device. The feature detection device according to claim 2,
The vocabulary learning curve consists of the data series of the ages arranged in the order of the cumulative number,
The smoothing unit smoothes the date data series using a Kalman smoother,
The detecting unit minimizes the vocabulary learning speed from a series of days required to learn one word obtained when smoothing, where the number of days required to learn one word is maximized. Detect as part
Feature detection device. A feature detection method in a feature detection apparatus including a first plateau detection unit and a second plateau detection unit,
Based on the vocabulary learning curve indicating the relationship between the age at which the infant started speaking a new word and the cumulative number of words that the infant started speaking before the age by the first plateau detection unit A first plateau detection step for detecting a portion where the interval between days of age when a new word is spoken is greater than a threshold;
Smoothing the vocabulary learning curve by a second plateau detection unit, and detecting a portion where the vocabulary learning speed of the smoothed vocabulary learning curve is minimized.
Feature detection method.   The program for functioning a computer as each part which comprises the characteristic detection apparatus in any one of Claims 1-3.

Images