Information processing device, information processing method, and information processing program

Abstract

This information processing device 1 is equipped with a calculation unit 13 for calculating the degree of overlap between the probability distribution of each probability value included in the i row or i column of a semantic similarity matrix and the probability distribution of each probability value included in the j row or j column (j≠i) of a waveform similarity matrix, as the degree of serendipity between the i-th word and the j-th word, by using said semantic similarity matrix, which uses, as elements thereof, the probability values in one row or in one column of the semantic similarity between words for a plurality of words, and a waveform similarity matrix which uses, as elements thereof, the probability values in one row or in one column of the waveform similarity between time-series data for time series data related to said words.

Description

[Technical field]

[0001] The present invention relates to an information processing device, an information processing method, and an information processing program. [Background technology]

[0002] With the recent trend towards big data, it is expected that new value will be created through data analysis. For example, if a large amount of data is collected from point of sales (POS) information, and evidence containing unexpected relationships that could not be found before is discovered from that data, it can be used for advanced market forecasting and sales strategy planning.

[0003] However, discovering the above evidence is not easy, and there is a possibility that unexpected relationships hidden in the data may be overlooked. Let us assume that there is a word pair consisting of two words, and two time series data corresponding to the two words are available. The semantic distance between two words is a subjective feeling that humans have, and it is expected that words that are semantically distant from each other are less related, and that the movements of the time series data will not be similar. For example, ham and cars are different types, so it is expected that their price movements will not be similar.

[0004] In this case, if the actual time series data (price movements in this example) is similar contrary to expectations, there is a surprise there. In other words, it can be considered that there is a surprise between two words that are not similar in meaning but have similar waveforms of time series data. Therefore, it is possible to define an index that represents the semantic similarity between two words and an index that represents the similarity of the waveforms, and define the index that combines these two indexes as the surprise between the two words.

[0005] For example, in Non-Patent Document 1, for two words i and j, the cosine similarity between the vector of word i and the vector of word j is called the semantic similarity u i,j The waveform similarity between the time series data related to word i and the time series data related to word j is expressed as the waveform similarity v i,j As the semantic similarity ui,j and the waveform similarity v i,j Distance d between i,j (=w v v i,j -w u u i,j ) is calculated as the degree of surprise. v And w u is a weighting factor.

[0006] Other index synthesis techniques include a keyword-based basket analysis of the occurrence frequency of specified data (Non-Patent Document 2) and a method of analyzing the main components of multiple variables that are correlated with each other (Non-Patent Document 3). However, the method in Non-Patent Document 2 is a method of analyzing co-occurrence, such as similar purchase frequencies of products, and cannot quantitatively indicate the relationship between heterogeneous data such as meaning and time series data. The method in Non-Patent Document 3 is based on the premise that data are homogeneous and correlated, and therefore cannot quantify the relationship between heterogeneous data such as meaning and time series data. [Prior art documents] [Non-patent literature]

[0007] [Non-Patent Document 1] Moriya, et al., "A Study on the Gap between Subjective and Objective Similarity of Words," Institute of Electronics, Information and Communication Engineers, 2020 Society Conference, A-10-12, September 2020. [Non-Patent Document 2] Motoda and 3 others, "Basics of Data Mining", Ohmsha, March 2008, p.41- p.43 [Non-Patent Document 3] Okuda and three others, “Multivariate analysis method”, Japan Society of Science and Technology, August 1986, p.159- p.163 Summary of the Invention [Problem to be solved by the invention]

[0008] In order to calculate the degree of unexpectedness between words, it is necessary to combine heterogeneous indices, namely, semantic and time-series data, as in Non-Patent Document 1. However, in Non-Patent Document 1, heterogeneous indices are directly calculated together to create a new composite index, which poses the following problems.

[0009] The first problem is that it is easily influenced by certain indicators. In Non-Patent Document 1, the waveform similarity between time series data v i,j Since the waveform similarity v is calculated using DTW (Dynamic Time Warping), i,j The range of values ​​of d becomes very large. i,j Substituting this into the compound equation, the distance d i,j The value of is the waveform similarity v i,j The semantic similarity u i,j and the waveform similarity v i,j The disadvantage is that the weights w u , w v Adjustment also requires trial and error.

[0010] Second, it is assumed that the index is accurate. It is difficult to quantify the similarity of waveforms in time series data, and even if it is calculated using DTW, the waveform similarity v i,j The waveform similarity v calculated using DTW cannot be quantified accurately. i,j If the value of is inaccurate, the distance d i,j In addition, it is difficult to quantify the scale by which a person perceives "waveforms are similar" when looking at them.

[0011] The third problem is that the scales of the indices are different. Combining two indices that are different in nature and have different scales on the same level, that is, combining indices with different scales for meaning and time series data, can lead to erroneous results. For example, adding or subtracting the raw values ​​of height and weight is meaningless.

[0012] The present invention has been made in consideration of the above circumstances, and an object of the present invention is to provide a technique capable of appropriately synthesizing a plurality of heterogeneous indices. [Means for solving the problem]

[0013] An information processing device of one embodiment of the present invention includes a calculation unit that uses a semantic similarity matrix whose elements are probability values ​​within a row or column of semantic similarity between a plurality of words, and a waveform similarity matrix whose elements are probability values ​​within a row or column of waveform similarity between time series data related to the words, to calculate the degree of overlap between the probability distribution of each probability value included in the i-th row or i-th column of the semantic similarity matrix and the probability distribution of each probability value included in the j-th row or j-th column (j ≠ i) of the waveform similarity matrix as a degree of surprise between the i-th word and the j-th word.

[0014] An information processing device of one embodiment of the present invention includes a calculation unit that uses a semantic similarity matrix whose elements are semantic similarities between a plurality of words and a waveform similarity matrix whose elements are waveform similarities between time series data related to the words, and calculates a composite value of the standardized variable value of the semantic similarity in the i-th row and j-th column of the semantic similarity matrix and the standardized variable value of the waveform similarity in the i-th row and j-th column of the waveform similarity matrix as a degree of surprise between the i-th word and the j-th word when the multiple semantic similarities included in the semantic similarity matrix and the multiple waveform similarities included in the waveform similarity matrix each follow a normal distribution or a Poisson distribution.

[0015] An information processing method of one embodiment of the present invention is an information processing method performed by an information processing device, which includes a step of using a semantic similarity matrix whose elements are probability values ​​within a row or column of semantic similarity between a plurality of words, and a waveform similarity matrix whose elements are probability values ​​within a row or column of waveform similarity between time series data related to the words, to calculate the degree of overlap between the probability distribution of each probability value included in the i-th row or i-th column of the semantic similarity matrix and the probability distribution of each probability value included in the j-th row or j-th column (j ≠ i) of the waveform similarity matrix as a degree of surprise between the i-th word and the j-th word.

[0016] An information processing method of one embodiment of the present invention is an information processing method performed by an information processing device, which includes a step of using a semantic similarity matrix whose elements are the semantic similarities between a plurality of words, and a waveform similarity matrix whose elements are the waveform similarities between time series data related to the words, and when the plurality of semantic similarities contained in the semantic similarity matrix and the plurality of waveform similarities contained in the waveform similarity matrix each follow a normal distribution or a Poisson distribution, calculating a composite value of the standardized variable value of the semantic similarity in the i-th row and j-th column of the semantic similarity matrix and the standardized variable value of the waveform similarity in the i-th row and j-th column of the waveform similarity matrix as a degree of surprise between the i-th word and the j-th word.

[0017] An information processing program according to one aspect of the present invention causes a computer to function as the information processing device. Effect of the Invention

[0018] According to the present invention, a technique capable of appropriately combining a plurality of different indices can be provided. [Brief description of the drawings]

[0019] [Figure 1] FIG. 1 is a reference diagram for explaining the outline of the present invention. [Diagram 2] FIG. 2 is a diagram showing a functional block configuration of the information processing device according to the first embodiment. [Diagram 3] FIG. 3 is a diagram showing a process flow for calculating the degree of surprise according to the first embodiment. [Figure 4] FIG. 4 is a reference diagram for explaining the calculation process flow of FIG. [Diagram 5] FIG. 5 is a diagram showing a conversion process flow for converting raw values ​​of a similarity matrix into probability values. [Figure 6] FIG. 6 is a reference diagram for explaining the conversion process flow of FIG. [Figure 7] FIG. 7 is a reference diagram for explaining the effects of the first embodiment. [Figure 8]FIG. 8 is a diagram showing a functional block configuration of an information processing device according to the second embodiment. [Figure 9] FIG. 9 is a diagram showing a flow of a process for calculating a degree of surprise according to the second embodiment. [Figure 10] FIG. 10 is a reference diagram for explaining the calculation process flow of FIG. [Figure 11] FIG. 11 is a reference diagram for combining the three indices. [Figure 12] FIG. 12 is a reference diagram for explaining the third embodiment. [Figure 13] FIG. 13 is a reference diagram for explaining the third embodiment. [Figure 14] FIG. 14 is a diagram illustrating a hardware configuration of an information processing device. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0020] Hereinafter, an embodiment of the present invention will be described with reference to the drawings. In the description of the drawings, the same parts are given the same reference numerals and the description will be omitted.

[0021] [1. Summary of the Invention] In the present invention, two methods are disclosed to solve the above problems.

[0022] Here, the semantic similarity u between the i-th and j-th word pair (i,j) i,j The semantic similarity matrix U has elements, and the waveform similarity v of the word pair (i,j) is i,j Let us assume that there is a waveform similarity matrix V whose elements are i,j is the raw value of the cosine similarity between the vector of word i and the vector of word j. The raw value is a number (e.g., 0.2) expressed as the distributed representation of words obtained by analyzing documents, etc. with word2vec, and is a value that has been subjected to general preprocessing such as outlier removal and logarithmic conversion. i,jis a raw value calculated using DTW to express the waveform similarity between time series data related to word i and time series data related to word j, such as the waveform similarity of time series data related to price fluctuations. Instead of DTW, a value that indexes the degree of similarity between two time series data using a well-known method, such as a correlation coefficient, may be used. In this case, the degree of surprise between rice and cucumber is calculated as follows:

[0023] The first method is to calculate the degree of surprise based on the shape of each distribution of rice and cucumber captured by row or column. First, the semantic similarity u i,j The raw values ​​of are converted into probability values ​​within a row or column, and the waveform similarity v included in the waveform similarity matrix V is calculated. i,j The raw values ​​of are also converted to probability values. Next, the semantic similarity of the converted probability values, u' i,j The semantic similarity matrix U' with elements and the waveform similarity v' of the transformed probability values i,j Using the waveform similarity matrix V' whose elements are x, y, and z, the degree to which the shape of the probability distribution for the rice row in the semantic similarity matrix U' is similar to the shape of the probability distribution for the cucumber row in the waveform similarity matrix V' (the degree of overlap of the distributions) is taken as the degree of surprise between rice and cucumber (see Figure 1(a)).

[0024] The second method is to calculate the unexpectedness based on the standardized variable values ​​of elements in a normal distribution or a Poisson distribution. First, the semantic similarity u i,j Assuming that the raw values ​​of follow a normal distribution, the waveform similarity v contained in the waveform similarity matrix V is i,j We assume that the raw values ​​of also follow a normal distribution. Next, the standardized variable values ​​Zu of rice and cucumber in the semantic similarity matrix U are 米,キュウリ Similarly, the standardized variable values ​​Zv 米,キュウリ Then, the standardized variable value Zu 米,キュウリ and the standardized variable value Zv 米,キュウリ The composite value of these two is taken as the degree of surprise between rice and cucumber (see Figure 1(b)).

[0025] The first and second methods use distribution information (standardized variable values ​​assumed to follow a probability distribution, normal distribution, or Poisson distribution) of each element of the semantic similarity matrix U and the waveform similarity matrix V when synthesizing heterogeneous indices, i.e., meaning and time series data, and can suppress the phenomenon in which the influence of either the semantic similarity or the waveform similarity becomes too strong. As a result, it is possible to provide a technology that can appropriately synthesize multiple heterogeneous indices.

[0026] [2. First embodiment] In the first embodiment, a first method will be described.

[0027] [2.1. Configuration of information processing device] 2 is a diagram showing a functional block configuration of an information processing device 1 according to a first embodiment. The information processing device 1 is an index synthesis device that synthesizes an index representing the semantic similarity of words and an index representing the waveform similarity of time-series data related to the words, and is a surprise degree calculation device that calculates the index value after the synthesis as a surprise degree. The information processing device 1 includes an acquisition unit 11, a conversion unit 12, and a calculation unit 13.

[0028] The acquisition unit 11 acquires the semantic similarity u between the i-th and j-th word pairs (i, j) read from the storage unit of the information processing device 1, the Internet, or the like, or input by the user. i,j The acquisition unit 11 is a functional unit for acquiring a semantic similarity matrix U having elements of the word pair (i, j). i,j This is a functional section that acquires a waveform similarity matrix V having the elements:

[0029] The conversion unit 12 converts all semantic similarities u i,j The raw values ​​of are converted into probability values ​​within a row or column, and the semantic similarity of the probability values ​​is calculated as u' i,j The conversion unit 12 is a functional unit that generates a semantic similarity matrix U′ having elements of all waveform similarities v i,j The raw values ​​of are converted into probability values ​​within a row or column, and the waveform similarity of the probability values ​​is calculated as v' i,j This is a functional section that generates a waveform similarity matrix V' having the elements:

[0030] The calculation unit 13 is a functional unit that uses the semantic similarity matrix U' and the waveform similarity matrix V' to calculate the degree of overlap between the probability distribution of each probability value included in the i row or i column of the semantic similarity matrix U' and the probability distribution of each probability value included in the j row or j column (j ≠ i) of the waveform similarity matrix V' as the degree of surprise between the i-th word and the j-th word.

[0031] [2.2. Operation of information processing device] Fig. 3 is a diagram showing a calculation process flow of the degree of surprise according to the first embodiment, and Fig. 4 is a reference diagram for explaining the calculation process flow.

[0032] Step S101; First, the acquisition unit 11 acquires the semantic similarity u i,j As mentioned above, we obtain the semantic similarity matrix U with the elements u i,j is the raw value of the cosine similarity between the vector of word i and the vector of word j. The raw value is a number (e.g., 0.2) expressed as the distributed representation of words obtained by analyzing documents, etc. with word2vec, and is a value that has been subjected to general preprocessing such as outlier removal and logarithmic conversion.

[0033] Step S102; Next, the conversion unit 12 converts all the semantic similarities u i,j The raw values ​​of are converted into probability values, with each row being set to 1, and the semantic similarity of the probability values ​​is calculated as u' i,j A semantic similarity matrix U′ having elements of the above is generated. The method of converting the semantic similarity matrix into a probability value will be described later.

[0034] The information processing device 1 calculates the waveform similarity v i,j Steps S101 and S102 are also executed for the waveform similarity matrix V having elements v' i,j A waveform similarity matrix V' having the elements

[0035] Step S103; Finally, the calculation unit 13 calculates, for example, the semantic similarity u' of the rice row from the semantic similarity matrix U'. 米,j (1≦j≦m) and calculate the semantic similarity u' of the relevant US bank. 米,j Similarly, the calculation unit 13 obtains the probability value of each of the waveform similarity v' of the cucumber row from the waveform similarity matrix V'. キュウリ,j (1≦j≦m) is taken out, and the waveform similarity v' of the cucumber row is calculated. キュウリ,j Then, the calculation unit 13 calculates the semantic similarity u' 米,j and the waveform similarity v' of the cucumber row キュウリ,j The degree of overlap with the probability distribution of is quantified using the Kullback-Leibler divergence formula (1), and the quantified value is the degree of surprise between rice and cucumber. 米,j and the waveform similarity v' of the cucumber row キュウリ,j If you plot these on a single radar chart, the degree of overlap between the two will be r 米,キュウリ Finally, the calculation unit 13 calculates the degree of surprise r i,j We obtain the surprisingness matrix R whose elements are

[0036]

number

[0037] Equation (1) expresses the semantic similarity u' of the rice row. 米,j and the waveform similarity v' of the cucumber row キュウリ,j It is a formula for indexing the degree of similarity with the distribution of . It is like calculating the similarity between the semantic vector of rice, calculated relatively within the entire vocabulary, and the waveform vector of cucumber, calculated relatively within the entire vocabulary. The smaller the value of D, the more similar the distributions are. The value of D is smallest when u'=v'. It quantifies the degree of overlap between the "semantic trend" and the "waveform trend" within the entire vocabulary. The closer the two are, the more the distribution shapes overlap, indicating that the meaning of rice and the waveform trend of cucumber are similar.

[0038] In step S103, D(V' キュウリ ||U' 米) was explained, but D(U' キュウリ ||V' 米 ), D(V' 米 ||U' キュウリ ), D(U' 米 ||V' キュウリ ), a similar effect can be obtained. The probability value may also be calculated column by column, with the entire column being set to 1.

[0039] [2.3. Conversion process] The method of conversion to a probability value performed in step S102 above will be described. As a method of conversion to a probability value, for example, a method using relative frequency as follows is considered. Any other method may also be used.

[0040] Fig. 5 is a diagram showing a conversion process flow for converting raw values ​​of a similarity matrix into probability values, and Fig. 6 is a reference diagram for explaining the conversion process flow.

[0041] Step S102a; First, the conversion unit 12 converts the semantic similarity u キュウリ,j The raw values ​​of are extracted, and a histogram is generated for the cucumber row in question, with the horizontal axis representing the class of the raw values ​​and the vertical axis representing the frequency of the raw values.

[0042] Step S102b; Next, the conversion unit 12 uses the histogram to calculate the semantic similarity u キュウリ,j Calculate the probability value of each of (1≦j≦m). For example, the semantic similarity u キュウリ,紙 belongs to the class of interval k, and the interval k has a frequency c(k), then the value of [c(k) / {c(1)+…+c(k)+c(N)}] is taken as the semantic similarity u キュウリ,紙 Let the probability value be .

[0043] The information processing device 1 also executes steps S102a and S102b for rows included in the semantic similarity matrix U other than the cucumber row.

[0044] Step S102c; Finally, the conversion unit 12 converts the semantic similarity u キュウリ,j The probability values ​​of all rows of are converted into an m × m matrix to obtain a semantic similarity matrix U' of the probability values.

[0045] Waveform similarity v i,j The raw values ​​of are also converted to probability values ​​using a similar procedure.

[0046] [2.4. Effects of the First Embodiment] In the first embodiment, the semantic similarity u i,j and the waveform similarity v i,j Since the same scale, the probability distribution, is used in both methods, the influence on specific indicators can be suppressed. In addition, since no weight adjustment parameters are used, weight adjustment is unnecessary.

[0047] In the first embodiment, the semantic similarity u i,j and the waveform similarity v i,j Since each element of is converted to a probability value, the semantic similarity u i,j and the waveform similarity v i,j This narrows the range of possible values ​​of and reduces the value of outliers, eliminating the assumption that the indicator is accurate.

[0048] In the first embodiment, the effect of individual values ​​is reduced because the data is captured on a line-by-line basis. In this regard, in Non-Patent Document 1, the word vector of word i and the word vector of word j are paired (first-stage relativization) to obtain the semantic similarity u i,j The waveform similarity v obtained by pairing the time series data of word i with the time series data of word j (first stage relativization) i,j In the first embodiment, the degree of surprise is calculated based on the semantic similarity u (relativized at the second stage) summarized by row or column, and the waveform similarity v (relativized at the second stage) summarized by row or column (see the present embodiment in FIG. 7). Since further relativization is performed compared to Non-Patent Document 1, the semantic similarity u i,j and waveform similarity v i,jThis has the effect of reducing the reliance on and requirement for accuracy, thereby eliminating the assumption that indicators are accurate.

[0049] In the first embodiment, the semantic similarity u i,j and the waveform similarity v i,j Since each element of is converted into a probability value, it can be properly combined even if the indices have different scales.

[0050] These features provide a technology that can appropriately combine multiple heterogeneous indices. Users can feel at ease because an accurate degree of surprise is automatically calculated even if they do not know the theoretical background. The degree of surprise for humans is difficult to evaluate because it is a matter of intuition, but by using the surprise algorithm described in the first embodiment, the results can be compared and used for calculation verification. In the future, this will also lead to the evaluation of the relative merits of algorithms by majority vote.

[0051] [3. Second embodiment] In the second embodiment, the second method will be described. As mentioned at the beginning, in the second embodiment, it is assumed that each element (i, j) of the semantic similarity matrix U and the waveform similarity matrix V follows a normal distribution or a Poisson distribution, and the mean value and variance value of the normal distribution or the Poisson distribution are explicitly given or automatically calculated from a sample, converted into a standardized variable value Z of a standard normal distribution, synthesized, or converted back into a raw value, and used as the unexpectedness degree of the word pair (i, j).

[0052] [3.1. Configuration of information processing device] 8 is a diagram showing a functional block configuration of an information processing device 1 according to the second embodiment. The information processing device 1 is also an index synthesis device and a surprise degree calculation device. The information processing device 1 includes an acquisition unit 21, a calculation unit 22, a determination unit 23, a conversion unit 24, a synthesis unit 25, an inverse conversion unit 26, and a calculation unit 27.

[0053] The acquisition unit 21 acquires the semantic similarity u between the i-th and j-th word pairs (i, j) read from the storage unit of the information processing device 1, the Internet, or the like, or input by the user. i,jThe acquisition unit 21 is a functional unit for acquiring a semantic similarity matrix U having elements v i,j This is a functional section that acquires a waveform similarity matrix V having the elements:

[0054] The calculation unit 22 calculates the semantic similarity u in the semantic similarity matrix U. i,j The calculation unit 22 is a functional unit that calculates a histogram of the waveform similarity v in the waveform similarity matrix V. i,j This is a functional section that calculates a histogram of

[0055] The determination unit 23 determines the semantic similarity u i,j The histogram of the semantic similarity u i,j The determining unit 23 is a functional unit that inquires of the user whether the waveform similarity v i,j The histogram of is displayed on the user terminal 3, and the waveform similarity v i,j This is a function that queries the user as to whether or not the semantic similarity u follows a normal distribution or a Poisson distribution. i,j and waveform similarity v i,j When it is assumed that follows a normal distribution or a Poisson distribution, this functional unit may be omitted.

[0056] In addition, the determination unit 23 determines the semantic similarity u i,j If follows a normal or Poisson distribution, the mean value μ of the normal or Poisson distribution of the semantic similarity matrix U u and the variance value σ u The determining unit 23 is a functional unit that determines the waveform similarity v i,j If follows a normal distribution or Poisson distribution, the mean value μ of the normal distribution or Poisson distribution related to the waveform similarity matrix V v and the variance value σ v This is a functional unit that determines the

[0057] The determination unit 23 also determines the average value μ u and the variance value σ uand the average value μ v and the variance value σ v The determining unit 23 is a functional unit that uses the semantic similarity matrix U and the waveform similarity matrix V to calculate the respective average values ​​and variance values ​​and determines them as the average value and variance values ​​to be used.

[0058] The conversion unit 24 converts the average value μ u and the variance value σ u Using the semantic similarity matrix U, i,j The raw values ​​of the standardized variable Zu i,j The conversion unit 24 is a functional unit that converts the average value μ v and the variance value σ v Using the waveform similarity matrix V, the waveform similarity v i,j The raw values ​​of the variables are standardized as Zv i,j This is the functional part that converts it into

[0059] The synthesis unit 25 calculates the semantic similarity u i,j The standardized variable value Zu i,j and the waveform similarity v i,j The standardized variable value Zv i,j This is a functional unit that synthesizes the above.

[0060] The inverse transformation unit 26 calculates the average value μ u and the variance value σ u and the average value μ v and the variance value σ v Using the standardized variable value (mean 0, variance 1), the composite value Zr i,j Let us use the unexpectedness r, which is a non-standardized variable value (mean not 1, variance not 1). i,j This is a functional part that converts the signal back into a raw value.

[0061] The calculation unit 27 calculates a plurality of semantic similarities u included in the semantic similarity matrix U by using the semantic similarity matrix U and the waveform similarity matrix V. i,j and multiple waveform similarities v included in the waveform similarity matrix V i,jIf and follow normal or Poisson distributions, respectively, the standardized variable value Zu of the semantic similarity in the i-th row and j-th column of the semantic similarity matrix U is i,j and the waveform similarity v of the i-th row and j-th column of the waveform similarity matrix V i,j The standardized variable value Zv i,j The composite value of Zr i,j , or the degree of surprise r obtained by the inverse conversion unit 26 i,j is a function that calculates the degree of surprise between the i-th word and the j-th word.

[0062] [3.2. Operation of information processing device] Fig. 9 is a diagram showing a calculation process flow of the degree of surprise according to the second embodiment, and Fig. 10 is a reference diagram for explaining the calculation process flow.

[0063] Step S201; First, the acquisition unit 11 acquires the semantic similarity u i,j As mentioned above, we obtain the semantic similarity matrix U with the elements u i,j is the raw value of the cosine similarity between the vector of word i and the vector of word j. The raw value is a number (e.g., 0.2) expressed as the distributed representation of words obtained by analyzing documents, etc. with word2vec, and is a value that has been subjected to general preprocessing such as outlier removal and logarithmic conversion.

[0064] Step S202; Next, the calculation unit 22 calculates the semantic similarity u i,j A histogram is calculated for (1≦i≦m, 1≦j≦m) with the horizontal axis representing the class of raw values ​​and the vertical axis representing the frequency of raw values. In the first embodiment, the histogram is considered row by row or column by column, but in the second embodiment, the histogram is considered for the entire matrix, and m×m items are used as the sum of the frequencies. In addition, if the matrix is ​​symmetric, for example, "waveform similarity between rice and cucumber" and "waveform similarity between cucumber and rice" will have the same value, and the diagonal of the matrix will be, for example, "waveform similarity between rice and rice", which will no longer make sense. Therefore, the number of elements obtained by excluding the symmetrical components (diagonal components) from the upper triangular matrix, or the number of elements obtained by excluding the symmetrical components from the lower triangular matrix, that is, mThe sum of the frequencies may be C2=m×(m-1)÷2, etc.

[0065] Then, the determination unit 23 determines the semantic similarity u i,j The histogram of the semantic similarity u i,j The user is asked whether follows a normal distribution (or a Poisson distribution). i,j If follows a normal distribution, proceed to the next step and calculate the semantic similarity u i,j If does not follow a normal distribution, the process ends without executing any further processing.

[0066] Step S203; The above semantic similarity u i,j If follows a normal distribution, the decision unit 23 selects the semantic similarity u as an external parameter. i,j The average value μ u and the variance value σ u The user is requested to input the average value μ u and the variance σ u Let us use the semantic similarity u i,j The average value μ u and the variance σ u It is determined as follows.

[0067] In this case, the semantic similarity u i,j Since the population of follows a normal distribution, the decision unit 23 determines the semantic similarity u i,j Using the histogram of the mean value μ u and the variance σ u may be calculated automatically. In that case, the average value μ u and the variance σ u is the semantic similarity u i,j The sample mean and sample variance are (1≦i≦m, 1≦j≦m).

[0068] Average value μ u and the variance σ u When inputting from outside, there is an advantage that these values ​​can be freely set. On the other hand, the average value μ u and the variance σ uWhen calculating automatically, the unexpectedness can be calculated automatically by simply inputting the semantic similarity matrix U and the waveform similarity matrix V, and thereafter, there is an advantage that there is no need to adjust the weights.

[0069] Step S204; Next, the conversion unit 24 converts the semantic similarity u i,j The average value μ u and the variance σ u Using the above, we calculate each semantic similarity u i,j For (1≦i≦m, 1≦j≦m), apply the conversion formula (2) to obtain the semantic similarity u i,j The standardized variable value Zu i,j get.

[0070]

number

[0071] Semantic similarity u i,j If the distribution of follows a log-normal distribution, the conversion unit 24 converts the semantic similarity u i,j It is also possible to apply the conversion formula (3) in which a logarithm is inserted between .

[0072]

number

[0073] The information processing device 1 calculates the waveform similarity v i,j Steps S201-S204 are also executed for the waveform similarity matrix V having elements. i,j The standardized variable value Zv i,j also gained.

[0074] Step S205; Next, the synthesis unit 25 applies the synthesis formula (4) to obtain the semantic similarity u i,j The standardized variable value Zu i,j and the waveform similarity v i,j The standardized variable value Zv i,j The composite value Zr i,j (1≦i≦m, 1≦j≦m).

[0075]

number

[0076] Step S206; Next, the inverse transform unit 26 converts each composite value Zr i,j For (1≦i≦m, 1≦j≦m), apply the inverse transformation formula of (5) to obtain the degree of surprise r i,j get.

[0077]

number

[0078] However, the parameter (σ r ,μ r ) is the semantic similarity u determined in step S203. i,j The average value μ u and the variance σ u and the waveform similarity v i,j The average value μ v and the variance σ v Using the above, the equation (6) for calculating the average value of the unexpectedness degree and the equation (7) for calculating the variance value of the unexpectedness degree are automatically calculated.

[0079]

number

[0080]

number

[0081] Step S207; Finally, the calculation unit 27 calculates the degree of surprise r i,j A surprise matrix R with elements (1≦i≦m, 1≦j≦m) is obtained.

[0082] The calculation unit 27 calculates the composite value Zr obtained in step S205 before applying the formula (5). i,jThe degree of surprise i,j Alternatively, the density function of the standard normal distribution before the inverse transformation can be expressed as "-∞ to synthetic value Zr i,j The integral value of " is the degree of surprise r i,j It is also possible to use the following.

[0083] [3.3. Modifications of the operation of the information processing device] Up to this point, an example has been described in which two indexes, an index of semantic similarity and an index of waveform similarity, are combined to calculate the degree of surprise. On the other hand, in the second embodiment, it is also possible to combine three or more indexes. Also, a normal distribution is obtained by adding or subtracting normal distributions. In particular, with regard to the index of waveform similarity, multiple types of waveform similarity are assumed, such as similarity regarding time series data of price fluctuations and similarity regarding time series data of stock price fluctuations.

[0084] Here, the semantic similarity u i,j The semantic similarity matrix U of the two and the waveform similarity v of the time series data of price fluctuations i,j The waveform similarity matrix V and the waveform similarity t i,j Assume that the waveform similarity matrix T of the above can be obtained. Also, a correspondence table between the price data of each item (i=rice, j=automobiles, k=PCs, ...) and the stock price data of each company (A=trading company, B=automobile manufacturer, C=electronics manufacturer, ...) is prepared in advance (see Figure 11(a)). Here, the price of rice i is considered to be linked to the stock price of trading company A, which sells rice, and the price of automobiles j is considered to be linked to the stock price of automobile manufacturer B.

[0085] In this case, the standardized variable value Zt i,j Replace i with A and j with B. This gives us three indicators: rice-related meaning, prices, and stock prices. We can then calculate the degree of surprise (R) by taking the difference between the "waveform similarity of prices and stock prices (V,T)" and the "semantic similarity (U)" (see Figure 11(b)).

[0086] Not limited to this, V, T, and U can be added, subtracted, multiplied, and divided. For example, since prices and stock prices are correlated, the standardized price variable value Zv i,j and the standardized variable value of stock price fluctuation Zt i,j Add and divide by 2 to obtain the standardized variable value Zu of the semantic similarity. i,j The composite value Zr i,j may be used as the degree of surprise.

[0087]

number

[0088] [3.4. Effects of the second embodiment] In the second embodiment, the semantic similarity u i,j and the waveform similarity v i,j Since the same scale, the standardized variable value Z, is used in both methods, the influence on a specific index can be suppressed. In addition, since no weight adjustment parameters are used, weight adjustment is not required.

[0089] In the second embodiment, the semantic similarity u i,j and the waveform similarity v i,j Since each element is expressed as the standardized variable value Z of the standard normal distribution, it can be appropriately synthesized even if the scales of the indices are different.

[0090] These features provide a technology that can appropriately combine multiple heterogeneous indices. Users can feel at ease because an accurate degree of surprise is automatically calculated, even if they do not know the theoretical background. The degree of surprise for humans is difficult to evaluate because it is a matter of intuition, but by using the surprise algorithm described in the second embodiment, the results can be compared and used for calculation verification. In the future, this will also lead to the evaluation of the relative merits and demerits of algorithms by majority vote.

[0091] [4. Third embodiment] [4.1. Importance of items to be combined] By carrying out the first or second embodiment, the degree of surprise r i,jA surprise matrix R with elements (1≦i≦m, 1≦j≦m) is obtained. However, when the number of items (m) increases explosively, the number of pairs becomes {m×(m-1)} / 2, so the more items there are, the more difficult it becomes for a person to view and interpret the surprise calculation results. Therefore, in the third embodiment, a method for calculating the importance of items to be combined with a given item will be described, in order to make it easier to select which item out of multiple items is important.

[0092] The information processing device 1 further includes an importance calculation unit that calculates the importance of an item in addition to the functional units shown in Fig. 2 or Fig. 8. The importance calculation unit calculates the importance of all the surprise degrees r i,j is converted into a probability value within a row or column, and the unexpectedness of the probability value r' i,j Next, the importance calculation unit calculates the surprise degree r' of the cucumber row from the surprise degree matrix R'. キュウリ,j Take out the unexpectedness r' キュウリ,j Each probability value is plotted on a radar chart (see Figure 12).

[0093] The importance calculator then applies equation (9) to quantify the peakiness of the probability distribution in the radar chart, which allows us to express in a scalar the likelihood that an item (cucumbers in this example) is dependent on a particular item.

[0094]

number

[0095] H is the entropy. キュウリ If is small, it suggests that cucumbers are biased toward some item and have a high value. Users can use this value to interpret it as "cucumbers are worth analyzing." For example, it can be used when formulating sales strategies for each item, such as which items should be purchased first.

[0096] [4.2. Importance of Similarity Matrix] The above method of calculating importance can be applied to grasp the importance of the similarity matrix in addition to grasping the importance of the target items of the combination. Since H represents uncertainty (peakedness), by comparing the "peakedness of waveform similarity" with the "peakedness of semantic similarity", it is possible to grasp which of the meanings and the waveforms is more significant.

[0097] The importance calculation unit calculates the semantic similarity u' of the cucumber row from the semantic similarity matrix U'. キュウリ,j and the semantic similarity u' キュウリ,j The importance calculation unit plots each probability value of the waveform similarity v' of the cucumber row from the waveform similarity matrix V' on a radar chart (see FIG. 13). キュウリ,j Then, the waveform similarity v' キュウリ,j The probability values ​​of are plotted on a radar chart. The importance calculation unit then applies formulas (10) and (11) to quantify the peakiness of the shape of the probability distribution in each radar chart.

[0098]

number

[0099]

number

[0100] As shown in Figure 13, HU' キュウリ HV' is not dependent on a specific item. キュウリ In this case, for cucumber, semantic similarity is less important, but waveform similarity is more important. Thus, HU' キュウリ Value of HV' キュウリ By comparing the values ​​of U and V, the importance of the semantic similarity matrix U and the waveform similarity matrix V can be considered.

[0101] [5.Other] The present invention is not limited to the above-described embodiment, and various modifications are possible within the scope of the present invention.

[0102] The information processing device 1 of the present embodiment described above can be realized, for example, by using a general-purpose computer system including a CPU 901, a memory 902, a storage 903, a communication device 904, an input device 905, and an output device 906, as shown in Fig. 14. The memory 902 and the storage 903 are storage devices. In the computer system, the CPU 901 executes a predetermined program loaded on the memory 902, thereby realizing each function of the information processing device 1.

[0103] The information processing device 1 may be implemented by one computer. The information processing device 1 may be implemented by multiple computers. The information processing device 1 may be a virtual machine implemented in a computer. The program for the information processing device 1 may be stored in a computer-readable recording medium such as an HDD, SSD, USB memory, CD, or DVD. The program for the information processing device 1 may also be distributed via a communication network. [Explanation of symbols]

[0104] 1: Information processing device 11: Acquisition part 12: Conversion section 13: Calculation section 21: Acquisition Department 22: Calculation section 23: Decision section 24: Conversion section 25: Synthesis section 26: Inverse conversion section 27: Calculation section 901:CPU 902: Memory 903: Storage 904: Communication equipment 905: Input device 906: Output device

Claims

1. an acquisition unit that acquires a semantic similarity matrix having elements of semantic similarities between a plurality of words, and acquires a waveform similarity matrix having elements of waveform similarities between time series data related to the words; a conversion unit that converts each semantic similarity included in the semantic similarity matrix into a probability value when the sum of a plurality of semantic similarities included in one row or one column is taken as the whole, and converts each waveform similarity included in the waveform similarity matrix into a probability value when the sum of a plurality of waveform similarities included in one row or one column is taken as the whole; a calculation unit that uses the converted semantic similarity matrix and the converted waveform similarity matrix to calculate a degree of overlap between a probability distribution of each probability value included in the i-th row or the i-th column of the converted semantic similarity matrix and a probability distribution of each probability value included in the j-th row or the j-th column (j≠i) of the converted waveform similarity matrix as a degree of surprise between the word in the i-th row or the i-th column and the word in the j-th row or the j-th column; An information processing device comprising:

2. a calculation unit which uses a semantic similarity matrix having elements each representing a semantic similarity between a plurality of words and a waveform similarity matrix having elements each representing a waveform similarity between time-series data related to the words, and when a user has determined that a plurality of semantic similarities included in the semantic similarity matrix and a plurality of waveform similarities included in the waveform similarity matrix each follow a normal distribution or a Poisson distribution, calculates a composite value of a standardized variable value of the semantic similarity in row i, column j of the semantic similarity matrix and a standardized variable value of the waveform similarity in row i, column j of the waveform similarity matrix as a degree of unexpectedness between the word in row i and the word in row j; An information processing device comprising:

3. a determination unit that determines a mean value and a variance value of a normal distribution or a Poisson distribution related to the semantic similarity matrix, and determines a mean value and a variance value of a normal distribution or a Poisson distribution related to the waveform similarity matrix; a conversion unit that converts the semantic similarities included in the semantic similarity matrix into standardized variable values ​​using average values ​​and variance values ​​related to the semantic similarity matrix, and converts the waveform similarities included in the waveform similarity matrix into standardized variable values ​​using average values ​​and variance values ​​related to the waveform similarity matrix; a synthesis unit that synthesizes the standardized variable value of the semantic similarity and the standardized variable value of the waveform similarity; an inverse conversion unit that inversely converts the standardized variable values ​​after synthesis into non-standardized variable values ​​using the mean value and variance value of the semantic similarity matrix and the mean value and variance value of the waveform similarity matrix, The calculation unit is The information processing apparatus according to claim 2 , wherein the standardized variable value after the synthesis, which is the synthesis value, or the non-standardized variable value instead of the synthesis value is calculated as a degree of surprise.

4. The determination unit is 4. The information processing apparatus according to claim 3, wherein an externally input mean value and variance value related to the semantic similarity matrix and a mean value and variance value related to the waveform similarity matrix are determined as the mean value and variance value to be used.

5. The determination unit is 4. The information processing apparatus according to claim 3, wherein the semantic similarity matrix and the waveform similarity matrix are used to calculate respective average values ​​and variance values, and the average values ​​and variance values ​​are determined as the average values ​​and variance values ​​to be used.

6. An information processing method performed by an information processing device, obtaining a semantic similarity matrix having elements each representing a semantic similarity between a plurality of words, and obtaining a waveform similarity matrix having elements each representing a waveform similarity between time series data related to the words; converting each semantic similarity included in the semantic similarity matrix into a probability value when the sum of a plurality of semantic similarities included in one row or one column is taken as the whole, and converting each waveform similarity included in the waveform similarity matrix into a probability value when the sum of a plurality of waveform similarities included in one row or one column is taken as the whole; a step of calculating, using the converted semantic similarity matrix and the converted waveform similarity matrix, a degree of overlap between a probability distribution of each probability value included in the i-th row or the i-th column of the converted semantic similarity matrix and a probability distribution of each probability value included in the j-th row or the j-th column (j≠i) of the converted waveform similarity matrix as a degree of surprise between the word in the i-th row or the i-th column and the word in the j-th row or the j-th column; An information processing method for performing the above.

7. An information processing method performed by an information processing device, a step of calculating a degree of surprise between the word in the i-th row and the word in the j-th row by combining a semantic similarity matrix having elements each representing a semantic similarity between a plurality of words and a waveform similarity matrix having elements each representing a waveform similarity between time-series data related to the words, when a user has determined that the plurality of semantic similarities included in the semantic similarity matrix and the plurality of waveform similarities included in the waveform similarity matrix each follow a normal distribution or a Poisson distribution; An information processing method for performing the above.

8. An information processing program that causes a computer to function as the information processing device according to any one of claims 1 to 5.

Images