JP2020047209A - Ontology processing apparatus and ontology processing program - Google Patents

Abstract

To provide an ontology processing apparatus and an ontology processing program capable of aiding in addition of more suitable words to an ontology that has superordinate/subordinate concept relations.SOLUTION: An ontology processing apparatus according to an embodiment of the present disclosure includes a coordinate conversion unit, a distance calculation unit, and an addition candidate extraction unit. The coordinate conversion unit extracts a word vector from a word vector holding unit, and converts the extracted word vector into a conversion vector in a predetermined space where a cosine similarity can be approximated by a Euclidean distance. The distance calculation unit calculates the distance between a hyperplane determined from a result of a seed word fitting process and a word in the conversion vector obtained by the conversion by the coordinate conversion unit. The addition candidate extraction unit extracts, based on the distance calculated by the distance calculation unit, a word that is a candidate for addition to the ontology in an ontology storage unit from words held in the word vector holding unit.SELECTED DRAWING: Figure 1

Description

  The present disclosure relates to an ontology processing device and an ontology processing program.

  2. Description of the Related Art Conventionally, in language processing technology, research on machine translation and dialog understanding has been continued for many years. In recent years, research on advanced knowledge processing using semantic information (concept) of words has been performed. One such technology is ontology technology. An "ontology" is a type of dictionary, which systematically organizes the concepts of words. With respect to the generation of the ontology, for example, techniques disclosed in Patent Documents 1 and 2 are disclosed.

  Conventionally, Non-Patent Document 1 discloses a method of creating a word vector by providing a large number of documents and performing automatic learning. Although there was a technology for creating a vector expression from words before that, the technology described in Non-Patent Document 1 enables expression of a complex concept. For example, an operation such as “France” − “Paris” + “Tokyo” ≒ “Japan” can be performed with the created vector. This expresses "France"-"Paris" as "the country whose capital is the city". Thus, in Non-Patent Document 1, the direction of the vector has some “meaning”.

JP 2009-110513 A JP 2017-182168 A

"Distributed representations of words and phrases and their compositionality" Tomas Mikolov, Ilya Stuskever, Kai Chen, Greg Scord, Canada

  However, in the ontology generated by the technique of Patent Document 1, each word is the same, and an ontology having a superordinate concept / lower concept relationship cannot be created. Therefore, it is desirable to provide an ontology processing device and an ontology processing program that can support more appropriate word addition for an ontology having a higher concept / lower concept relationship.

  An ontology processing apparatus according to an embodiment of the present disclosure includes a word vector holding unit that holds a plurality of words, a plurality of word vectors corresponding to a plurality of words, and an ontology storage unit that stores an ontology. I have. The ontology processing device further includes a seed word acquisition unit, a coordinate conversion unit, a distance calculation unit, an additional candidate extraction unit, and an ontology editing unit. The seed word obtaining unit obtains, as a seed word, a word corresponding to a lower concept of the specified word from the ontology in the ontology storage unit. The coordinate conversion unit extracts a word vector from the word vector holding unit, and converts the extracted word vector into a conversion vector into a predetermined space in which cosine similarity can be approximated by a Euclidean distance. The distance calculation unit calculates the distance between the hyperplane obtained from the result of the fitting process of the seed word and the word of the conversion vector obtained by the conversion by the coordinate conversion unit. The additional candidate extraction unit extracts a word that is a candidate to be additionally registered in the ontology of the ontology storage unit from the words stored in the word vector storage unit based on the distance calculated by the distance calculation unit. The ontology editing unit adds a part or all of the words of the addition candidates extracted by the addition candidate extraction unit to the ontology in the ontology storage unit.

An ontology processing program according to an embodiment of the present disclosure includes a word vector holding unit that holds a plurality of words, a plurality of word vectors corresponding to a plurality of words, and an ontology storage unit that stores an ontology. Have the computer perform the following steps:
(A) Acquiring a word corresponding to a subordinate concept of a specified word as a seed word from the ontology of the ontology storage unit. (B) Extracting a word vector from the word vector holding unit and converting the extracted word vector into a seed word. Is converted into a conversion vector into a predetermined space that can be approximated by the Euclidean distance. (C) The hyperplane obtained from the result of the fitting process of the seed word and the word of the conversion vector obtained by the conversion Calculating the distance (D) Based on the distance obtained by the calculation, extracting a word that is a candidate to be additionally registered in the ontology of the ontology storage unit from the words stored in the word vector storage unit (E) extraction Adding some or all of the additional candidate words to the ontology in the ontology storage unit

  According to the ontology processing device and the ontology processing program as an embodiment of the present disclosure, it is possible to support more appropriate word addition to an ontology having a relationship between a higher concept and a lower concept.

1 is a block diagram illustrating a functional configuration of an ontology processing device according to an embodiment of the present disclosure. FIG. 2 is an explanatory diagram (conceptual diagram) illustrating a configuration example of an ontology before a word is added by the ontology processing device in FIG. 1. FIG. 2 is an explanatory diagram (conceptual diagram) showing a configuration example of an ontology after a word has been added by the ontology processing device of FIG. 1. FIG. 2 is an explanatory diagram showing a configuration example of a word vector of a seed word created by a word vector creating unit in FIG. 1. FIG. 2 is an explanatory diagram illustrating a result of a contribution ratio calculation performed by a distance calculation unit in FIG. 1. FIG. 3 is an explanatory diagram showing conversion coefficients for rotating so as to fit the distribution of each seed word acquired by the distance calculation unit in FIG. 1. FIG. 2 is an explanatory diagram illustrating a word vector created by a word vector creating unit in FIG. 1. FIG. 2 is an explanatory diagram showing parameters obtained by rotating coordinates of a word vector created by a word vector creating unit in FIG. 1. 2 is a flowchart illustrating the operation of the ontology processing device of FIG. 1. FIG. 7 is a block diagram illustrating a modification of the functional configuration of the ontology processing device of FIG. 1.

  Hereinafter, an ontology processing device 1 according to an embodiment of the present disclosure will be described in detail with reference to the drawings. The following description is a specific example of the present disclosure, and the present disclosure is not limited to the following embodiments.

[Constitution]
FIG. 1 is a block diagram showing a functional configuration of an ontology processing device 1 according to the present embodiment. The ontology processing apparatus 1 includes, for example, a control unit 1, a word vector creation unit 2, an input / output unit 3, an ontology storage unit 4, and a document creation unit 5.

  The ontology storage unit 4 stores an ontology formed by a plurality of words (concepts). The ontology stored in the ontology storage unit 4 has a configuration in which superordinate concepts and subordinate concepts between words (concepts) can be associated. The ontology stored in the ontology storage unit 4 is not limited to other specific data description formats as long as the upper concept and the lower concept can be associated with each other. can do.

  The document storage unit 5 stores a large amount of document data (for example, various types of document data files such as text data).

  The word vector creation unit 2 extracts a plurality of words from a large amount of document data in the document storage unit 5, and creates a plurality of word vectors (word vectors related to words included in the document data) corresponding to the extracted words. . The word vector creating unit 2 stores a plurality of extracted words and a plurality of word vectors corresponding to the plurality of extracted words. The word vector creation unit 2 may generate a word vector from the document data held in the document storage unit 5 using, for example, the method described in Non-Patent Document 1 or general word2vec. The method of generating a word vector from document data is not limited to the above method. When decomposing a document into words, the word vector creating unit 2 may refer to words stored in the ontology storage unit 4 together with a word dictionary (not shown) inside the word vector creating unit 2.

  The control unit 1 has a function of controlling each component of the ontology processing apparatus 1, and includes a processing target selection unit 11, a coordinate conversion unit 12, a distance calculation unit 13, and an ontology editing unit 14.

  The processing target selection unit 11 receives from the user a designation of a position (word) to search for an additional candidate on the ontology stored in the ontology storage unit 4. Hereinafter, the word specified by the user at this time is referred to as a basic word Wb. FIG. 2 illustrates “dog” as the basic word Wb. The processing target selection unit 11 is configured to extract a plurality of additional words (hereinafter, referred to as “peripheral words Wa”) based on the basic words Wb (concepts) specified by the user. To determine. FIG. 2 exemplifies a plurality of seed words Ws such as “Shiba Inu”, “Chihuahua”, “Maltes”, “Bulldog”, and “Dalmatian”. The processing target selection unit 11 acquires a word vector corresponding to the determined plurality of seed words Ws from the word vector creation unit 2.

  The coordinate conversion unit 12 extracts from the word vector creation unit 2 a word vector that can be approximated in a space in which the cosine similarity can be approximated by a Euclidean distance (hereinafter, referred to as “space A”). The coordinate conversion unit 12 performs coordinate conversion of the word vector extracted from the word vector creation unit 2 into the space A. Specifically, the coordinate conversion unit 12 first calculates the seed center from the average of the word vectors of the plurality of seed words Ws. Next, the coordinate conversion unit 12 converts a word vector corresponding to a word (seed word Ws, peripheral word Wa) whose cosine similarity with the seed center falls within a predetermined range (for example, 0.6 or more) into a word vector creation unit. Extract from 2. FIG. 3 illustrates “poodle” and “dachshund” as the peripheral words Wa.

  The coordinate conversion unit 12 normalizes the word vector corresponding to each seed word Ws extracted from the word vector creation unit 2 and normalizes the seed center (vector) becomes (1, 0,..., 0). Thus, each normalized word vector is rotated about the origin. The coordinate conversion unit 12 also normalizes the word vectors corresponding to the respective peripheral words Wa, and normalizes each of the normalized words such that the result (vector) obtained by normalizing the seed center is (1, 0,..., 0). Rotate the word vector about the origin. Here, when the coordinate conversion unit 12 converts the normalized seed center into polar coordinates (coordinates in the space A) described later, the normalized seed center becomes the origin in polar coordinates (coordinates in the space A). Then, each normalized word vector is rotated about the origin. Next, the coordinate conversion unit 12 converts the coordinates of each of the seed words Ws and each of the surrounding words Wa into polar coordinates (coordinates in the space A), and a vector (hereinafter referred to as “vector”) one-dimensionally smaller than the vector of the word vector creation unit 2. This is referred to as a "post-conversion vector." Hereinafter, the dimension of the transformed vector is represented as N dimensions (N = L-1). The distance calculation unit 13 treats the converted vector (the vector in the space A) approximately as a vector in the orthogonal coordinates.

Here, the cosine similarity is equivalent to the distance between two points on the hypersphere. The word vector is L-dimensional, and coordinates (y ₁ , y ₂ ,..., Y _L ) on the hypersphere are converted to polar coordinates to obtain (r, θ ₁ , θ ₂ ,..., Θ _L-1 ). At this time, since r is always 1, it is considered that a distance between two points is obtained using (θ ₁ , θ ₂ ,..., Θ _{L -1} ) excluding r. Because geometric on hypersphere, a non-Euclidean geometry, the Pythagorean theorem is not _{satisfied, (θ 1, θ 2,} ..., θ L-1) be calculated Euclidean distance is replaced in the orthogonal coordinates, on a spherical surface Does not generally coincide with the distance. However, if one point is fixed at (π / 2, π / 2,..., Π / 2, 0) and the other point is in the vicinity, it can be approximated by rectangular coordinates. (Π / 2, π / 2,..., Π / 2, 0) are coordinates obtained by converting (1, 0,..., 0) into polar coordinates and excluding r. In the present embodiment, a space (space A) in which the cosine similarity can be approximated by a Euclidean distance is used by using this method. When the word vector is L-dimensional, the space A has L-1 dimensions.

  The distance calculation unit 13 performs a process of searching for a candidate word to be added to the lower concept of the position (word) received by the processing target selection unit 11. Specifically, the distance calculation unit 13 first performs coordinate rotation using the seed center (the converted vector) as the origin so as to fit the distribution of the converted vectors (for example, see FIG. 4) of the plurality of seed words Ws. Do. The distance calculation unit 13 performs fitting processing of a plurality of seed words Ws (for example, fitting processing using calculation similar to principal component analysis). The distance calculation unit 13 determines an M-dimensional hyperplane based on the converted vector of the plurality of seed words Ws from the result of the fitting processing of the plurality of seed words Ws.

Hereinafter, a specific example of the determination of the M-dimensional hyperplane will be described. Assuming that the word vector creation unit 2 creates an N-dimensional vector, the converted vectors of a plurality of seed words Ws are represented, for example, as shown in FIG. At this time, the data in FIG. 4 is represented by a matrix shown in the following equation (1).

  The distance calculation unit 13 obtains a variance-covariance matrix of this matrix, and further obtains its eigenvalue and eigenvector. The distance calculation unit 13 arranges the eigenvectors in descending order of the eigenvalues to form a rotation matrix (FIG. 6). The distance calculation unit 13 calculates the contribution rate by dividing each eigenvalue by the sum of the eigenvalues (FIG. 5). The distance calculation unit 13 considers the cumulative contribution rate in FIG. 5 in order from the first component (PC1) and sets M as the dimension when it exceeds 90% for the first time. The distance calculation unit 13 sets a hyperplane extending from the first axis to the Mth axis as an M-dimensional hyperplane to be obtained. In FIG. 5, the dimension of the M-dimensional hyperplane is 3.

  Next, the distance calculation unit 13 performs coordinate rotation on the converted vector of the plurality of peripheral words Wa, and obtains the coordinates of the converted vector obtained by the coordinate rotation and the result of the fitting processing of the plurality of seed words Ws. Calculate the distance to the given hyperplane. Further, the distance calculation unit 13 extracts a word that is a candidate to be additionally registered in the ontology of the ontology storage unit 4 from the plurality of peripheral words Wa based on the calculated distance.

Hereinafter, a specific example of a method of extracting a candidate word to be additionally registered in the ontology of the ontology storage unit 4 will be described. FIG. 7 shows an example of a vector after conversion of a plurality of peripheral words Wa. The distance calculation unit 13 rotates the converted vectors of the plurality of peripheral words Wa using a rotation matrix as shown in FIG. At this time, the distance calculation unit 13 obtains the rotation result of the converted vector of the plurality of peripheral words Wa by, for example, the matrix operation shown in the following Expressions (2), (3), (4), and (5). (FIG. 8). In FIG. 8, the distance between the M-dimensional hyperplane and the j-th word is the distance from the hyperplane extending from the first axis to the Mth axis to the point indicated by the vector in FIG. The distance calculation unit 13 calculates the distance between the M-dimensional hyperplane and the j-th word by the following equation (6).

  The ontology editing unit 14 extracts the peripheral word Wa having a small distance between the M-dimensional hyperplane and the j-th word, and displays the word as an additional candidate word in the ontology together with the distance value via the output unit 32 (FIG. 3). . The ontology editing unit 14 receives a designation of a word to be added to the ontology in the ontology storage unit 4 among words of addition candidates from the user. Then, the ontology editing unit 14 performs a process of adding some or all of the addition candidate words to the ontology in the ontology storage unit 4 according to the user's instruction (operation).

  The input / output unit 3 has a function of a user interface, and includes an input unit 31 for receiving an operation and information input from a user, and an output unit 32 for outputting information to the user. As the input unit 31, for example, an input device such as a keyboard or a mouse can be applied. Further, as the output unit 32, an output device such as a display or a printer can be applied.

[motion]
Next, the operation of the ontology processing apparatus 1 according to this embodiment having the above configuration will be described. FIG. 9 is a flowchart illustrating the operation of the ontology processing device 1.

  In the flowchart of FIG. 9, as a premise (initial state), the creation of the word vector by the word vector creation unit 2 (the creation of the word vector using the document data in the document storage unit 5) is completed, and the created word vector is It is assumed that data is retained. Regarding the word vector creation processing by the word vector creation unit 2, as described above, the same processing as in Non-Patent Document 1 or general word2vec can be applied, and therefore a detailed description is omitted.

  In the flowchart of FIG. 9, it is assumed that an ontology composed of an arbitrary number of words (concepts) is registered in the ontology storage unit 4 as a premise (initial state).

  First, the processing target selection unit 11 receives designation (selection) of a position (basic word Wb) where a word of a lower concept is to be added on the ontology of the ontology storage unit 4 from the user (step S101). For example, the processing target selection unit 11 presents a word (concept) included in the ontology of the ontology storage unit 4 to the user via the input / output unit 3 (for example, displays a list or a map of words forming the ontology). Then, the designation (selection) of any word (concept) is accepted.

  Next, based on the concept (basic word Wb) specified in step S101, the processing target selection unit 11 acquires a seed word Ws as a seed when extracting an additional candidate word. Then, the processing target selection unit 11 acquires the word vector of the seed word Ws acquired in step S101 from the word vector creation unit 2 (step S102). When the word vectors corresponding to all the seed words Ws do not exist in the word vector creation unit 2, the processing target selection unit 11 acquires, for example, only the existing word vectors from the word vector creation unit 2.

  In this embodiment, the processing target selection unit 11 acquires a concept (word) one level lower than the concept specified by the user in step S101 from the ontology stored in the ontology storage unit 4. However, when the concept one level lower than the specified concept is an intermediate concept, the processing target selection unit 11 acquires (refers to) a concept one level lower than the intermediate concept as a lower level concept.

  Here, an example in which the concept (word) of “dog” is specified by the user from the concepts (words) stored in the ontology storage unit 4 will be described.

  FIG. 2 is an explanatory diagram illustrating an example in which the processing target selection unit 11 acquires the seed word Ws from the concept of “dog”. In the example of FIG. 2, the concepts “Shiba Inu”, “Pet Dog”, “Bulldog”, and “Dalmatian” exist as one concept below the dog. In the example of FIG. 2, “pet dog” is an intermediate concept, and an example is shown in which “Chihuahua” and “Maltese” exist as concepts one level below “pet dog”. Therefore, the processing target selection unit 11 acquires “Chihuahua” and “Maltese”, which are one level lower than the intermediate concept of “pet dog”, as a part of the seed word Ws corresponding to “dog”. Therefore, in the example of FIG. 2, the processing target selection unit 11 acquires “Shiba Inu”, “Chihuahua”, “Maltese”, “Bulldog”, and “Dalmatian” as seed words corresponding to the concept of “dog”. Then, the processing target selection unit 11 acquires the word vector of each seed word Ws for “dog” from the word vector creation unit 2.

  Next, the coordinate conversion unit 12 extracts a word vector that can be approximated in the space A from the word vector creation unit 2 and performs coordinate conversion to the space A (step S103). First, the coordinate conversion unit 12 calculates a seed center from an average of word vectors of a plurality of seed words Ws. Next, the coordinate conversion unit 12 extracts words (seed word Ws) and word vectors whose cosine similarity with the seed center is within a predetermined range (for example, 0.6 or more) from the word vector creation unit 2. The coordinate conversion unit 12 normalizes the word vector corresponding to each seed word Ws extracted from the word vector creation unit 2 and normalizes the seed center (vector) becomes (1, 0,..., 0). Thus, each normalized word vector is rotated about the origin. The coordinate conversion unit 12 also normalizes the word vectors corresponding to the respective peripheral words Wa, and normalizes each of the normalized words such that the result (vector) obtained by normalizing the seed center is (1, 0,..., 0). Rotate the word vector about the origin. Next, the coordinate conversion unit 12 converts the coordinates of each of the seed words Ws and each of the surrounding words Wa into polar coordinates (coordinates in the space A), and a vector (hereinafter referred to as “vector”) one-dimensionally smaller than the vector of the word vector creation unit 2. This is referred to as a "post-conversion vector." The distance calculation unit 13 treats the converted vector (the vector in the space A) approximately as a vector in the orthogonal coordinates.

  Next, the distance calculation unit 13 sets the seed center (the converted vector) as the origin so as to fit the distribution of each word vector (the word vector corresponding to each extracted seed word Ws) acquired in step S103. The coordinate rotation is performed, and an M-dimensional hyperplane based on each word vector is determined using the result (step S104).

  Hereinafter, a specific example of a process in which the distance calculation unit 13 determines an M-dimensional hyperplane based on each seed word Ws will be described.

  First, it is assumed that the word vector creating unit 2 has created an N-dimensional (N is an integer of 1 or more) vector as a word vector. Then, the word vector corresponding to the selected seed word Ws can be represented by a matrix (table) as shown in FIG. In the matrix (table) in FIG. 4, the seed word Ws is assigned to each row. In FIG. 4, “Shiba Inu”, “Chihuahua”, “Maltese”, “Bulldog”, and “Dalmatian” are assigned in order from the first line. In the matrix (table) of FIG. 4, parameters X1, X2, X3,..., XN are assigned in order from the first column.

  At this time, the data of FIG. 4 is represented by a matrix as in equation (1). In this case, the distance calculation unit 13 obtains a variance-covariance matrix of the matrix of Expression (1), and further obtains its eigenvalue and eigenvector. Then, the distance calculation unit 13 arranges the eigenvectors in descending order of the eigenvalues to obtain a rotation matrix (see FIG. 6). Then, the distance calculation unit 13 calculates the contribution rate by dividing each eigenvalue by the sum of the eigenvalues. Then, the distance calculation unit 13 cumulatively adds the contribution rates in order from the contribution rate of the eigenvalue having the largest eigenvalue, and calculates the cumulative contribution rate.

FIG. 5 shows the contribution ratio of each component (first component PC1, second component PC2,..., Nth component PCN) obtained as a result of processing the word vector of each seed word Ws, and the cumulative contribution ratio of each component (No. An example of a cumulative value (total value) of contribution rates from one component to the component is shown. FIG. 6 is a matrix showing the conversion coefficients corresponding to each combination of each parameter (X1 to XN) constituting each word vector and each component (PC1 to PCN). For example, it illustrates a transform coefficient transform coefficients corresponding to the combination of X1 and PC1 shown with a _11, corresponding to the combination of X and PC2 and a _12.

  Next, the distance calculation unit 13 refers to the cumulative contribution rate in order from the first component PC1 in the fitting result, and determines the number of dimensions M when the cumulative contribution rate T exceeds the predetermined cumulative contribution rate T for the first time (for the first time, the cumulative contribution rate T that is a threshold The number (order) of the component when it exceeds the number is acquired. Here, as an example, the cumulative contribution ratio T is assumed to be 90% (0.90), but the value of the cumulative contribution ratio T can be set arbitrarily.

  For example, in the result of the fitting shown in FIG. 5, when the cumulative contribution rate is referred to in order from the first component PC1, the third component PC3 has the cumulative contribution rate exceeding 90% (0.90) for the first time. Therefore, the distance calculation unit 13 acquires “3” as the number of dimensions M.

  Next, the distance calculation unit 13 determines a hyperplane extending from the first axis (the axis of the first component) to the Mth axis (the axis of the Mth component) as an M-dimensional hyperplane to be obtained. Here, since M = 3 as described above, the distance calculation unit 13 determines the hyperplane formed by the first axis, the second axis, and the third axis as the M-dimensional hyperplane to be obtained.

  As described above, the distance calculation unit 13 determines an M-dimensional hyperplane.

  Next, the distance calculation unit 13 calculates the distance between the M-dimensional hyperplane determined in step S104 and the word of each word vector held by the word vector creation unit 2 (step S105).

  Hereinafter, a specific example of the distance calculation processing between the M-dimensional hyperplane and each word vector performed by the distance calculation unit 13 will be described.

  FIG. 7 is an explanatory diagram showing an example of each word vector in the word vector creation unit 2. FIG. 7 illustrates each parameter (X1 to XN) of a word vector corresponding to each of the words “poodle”, “cockatiel”, “dachshund”,. In FIG. 7, the values of the parameters X1 to XN corresponding to the word "poodle" are shown as x11 to x1N. In FIG. 7, the values of the parameters X1 to XN corresponding to the word “cockatiel” are shown as x21 to x2N. Further, in FIG. 7, the values of the parameters X1 to XN corresponding to the word “Dachshund” are illustrated as x31 to x3N.

  FIG. 8 is an explanatory diagram showing a result of coordinate rotation of each word vector shown in FIG. Hereinafter, the result of the coordinate rotation of the word vector is referred to as a “rotation result vector”. In FIG. 8, word vectors corresponding to the words “poodle”, “cockatiel”, “dachshund”,... Are illustrated by parameters of the rotation result vector (first component PC1 to Nth component PCN).

8, the value of the first component PC1~ the N component PCN corresponding to the word "poodle" depicts a z ₁₁ to z _1N. In FIG. 8, the values of the first to Nth components PC1 to PCN corresponding to the word “cockatiel” are shown as z ₂₁ to z _2N . Further, in FIG. 8, the values of the first component PC1 to the Nth component PCN corresponding to the word “Dachshund” are illustrated as z ₃₁ to z _3N . Further, in FIG. 8, the values of the first component PC1 to the Nth component PCN corresponding to an arbitrary word are illustrated as z _i1 to z _iN (i is an integer of 1 or more). Note that the first component PC1~ value of the N component PCN corresponding to each word (z _i1 to z _iN), can be obtained by a matrix calculation shown in the equation (2). Note that in Expression (2), A is a matrix of Expression (3). In Expression (2), X is a matrix of Expression (4). Further, in Expression (2), Z is a matrix of Expression (5). Therefore, the distance calculation unit 13 can obtain the distance between the M-dimensional hyperplane and an arbitrary word (word vector) from the sum of squares of the parameter of the rotation result vector shown in FIG.

  Specifically, the distance calculation unit 13 can calculate the distance between the M-dimensional hyperplane and an arbitrary word (the i-th word) by Expression (6). For example, when M = 2, the distance D1 of the word “poodle” shown in FIG. 7 and FIG. 8 can be expressed as in equation (7).

  As described above, the distance calculation unit 13 calculates the distance between the M-dimensional hyperplane and each word vector (word vector other than the seed word) in the word vector creation unit 2.

  Next, the ontology editing unit 14 extracts words to be added to the ontology storage unit 4 from the words held by the word vector creation unit 2 based on the distance between the words calculated in step S105, and Is presented to the user (for example, output to the user via the input / output unit 3) (step S106).

  Next, the ontology editing unit 14 receives, from the user, whether or not addition is necessary for the addition candidate word (for example, accepts an input via the input / output unit 3), and, for the word input to add, adds the ontology storage unit. 4 is additionally registered in the ontology 4 (step S107).

  For example, the ontology editing unit 14 may extract a word whose distance is shorter (shorter) than a predetermined threshold as an addition candidate and present it to the user. For example, the ontology editing unit 14 displays an operation screen (GUI screen) via the input / output unit 3 to present words of addition candidates to the user, and determines whether each addition candidate word needs to be added to the ontology ( “Add” or “Do not add” may be accepted. In addition, the ontology editing unit 14 may add the information of the calculated distance to each additional candidate word as shown in FIG. 3 and present the word to the user.

  Next, the control unit 1 receives the necessity of the continuation of the process from the user (step S108), and starts the above-described step S101 when receiving the input of the continuation of the process from the user, and continues the process from the user. If an input indicating not to be received is received, the process ends.

[effect]
Next, effects of the ontology processing apparatus 1 according to the present embodiment will be described.

  In the ontology processing apparatus 1 according to the present embodiment, the word vector extracted from the word vector creation unit 2 is coordinate-transformed into a space A in which the cosine similarity can be approximated by the Euclidean distance. Furthermore, in the ontology processing apparatus 1 according to the present embodiment, the distance between the M-dimensional hyperplane determined based on the result of the fitting processing of the word vector of the seed word Ws and the word of the transformed vector coordinate-transformed to the space A Is calculated. As described above, in the ontology processing apparatus 1 according to the present embodiment, the word vector is coordinate-transformed into a space (space A) in which the cosine similarity can be approximated by the Euclidean distance, so that the conceptual relationship between words ( Synonymousness) is evaluated. In this way, the conceptual relationship between words (similarity) is more accurately compared to the case where the conceptual relationship (similarity) between words is evaluated simply by comparing the distances between word vectors. Can be evaluated. Therefore, it is possible to support more appropriate word addition for an ontology having a superordinate concept / subordinate concept relationship.

  In the ontology processing apparatus 1 according to the present embodiment, the seed center is calculated from the average of the word vectors of the plurality of seed words Ws, and the cosine similarity with the calculated seed center corresponds to a word within a predetermined range. The word vector to be extracted is extracted from the word vector creation unit 2, each word vector extracted from the word vector creation unit 2 is normalized, and each normalized normalized so that the normalized seed center becomes the origin in the space A. The word vector is converted to a conversion vector after being rotated about the origin. As a result, an error caused by the conversion into the space A can be minimized. Therefore, the conceptual relationship between words (similarity) can be accurately evaluated. Therefore, it is possible to support more appropriate word addition for an ontology having a superordinate concept / subordinate concept relationship.

  In the ontology processing apparatus 1 according to the present embodiment, if a certain number of words (concepts) are registered in the ontology of the ontology storage unit 4, words of additional candidates can be automatically extracted as seed words. .

  Further, in the ontology processing apparatus 1 according to the present embodiment, the distance from the M-dimensional hyperplane determined based on the result of the fitting processing of the word vector of the seed word is calculated, and the candidate is added based on the calculated distance. Words are being extracted. Thereby, the ontology processing apparatus 1 according to the present embodiment focuses on the similarity from a certain viewpoint, not on the overall similarity, and adds a new word to an ontology having an existing superordinate concept and subordinate concept. Can be.

  Further, in the ontology processing apparatus 1 according to the present embodiment, since the extracted addition candidate is displayed with the distance value together with the word (for example, see FIG. 4), it is finally added to the ontology according to the operation of the user. You can decide if you want to.

  In the ontology processing device 1 according to the present embodiment, the processing target selection unit 11 has specified the basic word Wb by the specification from the user. However, in the ontology processing apparatus 1 according to the present embodiment, the processing target selection unit 11 may specify a word defined in advance as the basic word Wb. Further, in the ontology processing apparatus 1 according to the present embodiment, the processing target selection unit 11 sequentially specifies the basic words Wb according to the specification from the user. However, in the ontology processing apparatus 1 according to the present embodiment, the processing target selection unit 11 may specify some or all of the words extracted by the word vector creation unit 2 as the basic words Wb.

  Further, in the ontology processing apparatus 1 according to the present embodiment, the distance calculation unit 13 sets the threshold of the cumulative contribution ratio when determining the dimension of the hyperplane to 90%. However, in the ontology processing apparatus 1 according to the present embodiment, the distance calculation unit 13 may set the threshold value of the cumulative contribution rate when determining the dimension of the hyperplane to a value different from 90%. At this time, the distance calculation unit 13 may allow the user to arbitrarily set the threshold value of the cumulative contribution rate when determining the dimension of the hyperplane.

  Further, in the ontology processing apparatus 1 according to the present embodiment, the index used when evaluating the conceptual relationship between words (similarity) is the distance between the M-dimensional hyperplane and the converted vector of the word. However, in the ontology processing apparatus 1 according to the present embodiment, an index used when evaluating the conceptual relationship between words (similarity) is a value other than the distance between the M-dimensional hyperplane and the vector after conversion of the word. It may be.

<4. Modification>
Hereinafter, a modified example will be described.

  FIG. 10 is a block diagram showing a modification of the functional configuration of the ontology processing device 1 according to the above embodiment. The ontology processing apparatus 1 according to the present modification is configured by configuring part or all of the ontology processing apparatus 1 according to the above-described embodiment as software. The ontology processing device 1 according to the present modification corresponds to, for example, the ontology processing device 1 according to the above-described embodiment in which the control unit 1 and the word vector creation unit 2 are replaced with an information processing unit 6 and a memory unit 7. I do. The memory unit 7 includes, for example, a hard disk drive or a flash memory, and stores an ontology processing program 71. The information processing section 6 performs the same processing as the processing executed by the control section 1 and the word vector creation section 2 by loading the ontology processing program 71 stored in the memory section 7.

  In this modification, the information processing unit 6 loaded with the ontology processing program 71 performs the same processing as the processing executed by the control unit 1 and the word vector creation unit 2. Therefore, the ontology processing device 1 according to the present modification has the same effects as those of the above embodiment.

  As described above, the present disclosure has been described with reference to the embodiment and the modifications thereof. However, the present disclosure is not limited to the above-described embodiment and the like, and various modifications are possible. Note that the effects described in this specification are merely examples. The effects of the present disclosure are not limited to the effects described in this specification. The present disclosure may have effects other than those described herein.

In addition, the present disclosure may have the following configurations.
(1)
A word vector holding unit that holds a plurality of words and a plurality of word vectors corresponding to the plurality of words,
An ontology storage unit that stores the ontology,
From the ontology of the ontology storage unit, a seed word acquisition unit that acquires a word corresponding to a lower concept of a specified word as a seed word,
A coordinate conversion unit that extracts the word vector from the word vector holding unit and converts the extracted word vector into a conversion vector into a predetermined space in which cosine similarity can be approximated by a Euclidean distance;
A hyperplane determined from the result of the fitting process of the seed word, and a distance calculation unit that calculates a distance between the word of the conversion vector obtained by the conversion in the coordinate conversion unit,
Based on the distance calculated by the distance calculation unit, from the words held in the word vector holding unit, an additional candidate extraction unit that extracts a word that is a candidate to be additionally registered in the ontology of the ontology storage unit,
An ontology editing unit for adding a part or all of the words of the addition candidate extracted by the addition candidate extraction unit to the ontology in the ontology storage unit.
(2)
The coordinate conversion unit calculates a seed center from an average of a plurality of word vectors of the seed word, and calculates a word vector corresponding to a word whose cosine similarity with the seed center obtained by the calculation is within a predetermined range. Normalizing each of the word vectors extracted from the word vector holding unit and each of the word vectors extracted from the word vector holding unit, such that the normalized seed center is the origin in the predetermined space; Is rotated around the origin and then converted into the conversion vector. The ontology processing apparatus according to (1).
(3)
A plurality of words, a word vector holding unit that holds a plurality of word vectors corresponding to the plurality of words, and a computer including an ontology storage unit that stores an ontology,
From the ontology in the ontology storage unit, acquiring a word corresponding to a lower concept of a specified word as a seed word;
Extracting the word vector from the word vector holding unit, and converting the extracted word vector into a conversion vector into a predetermined space in which the cosine similarity of the seed word can be approximated by a Euclidean distance;
Calculating the distance between the hyperplane determined from the result of the fitting process and the word of the conversion vector obtained by the conversion,
Extracting words that are candidates to be additionally registered in the ontology in the ontology storage unit, from the words stored in the word vector storage unit, based on the distance obtained by the calculation;
Adding, to the ontology in the ontology storage unit, a part or all of the extracted additional candidate words.

  DESCRIPTION OF SYMBOLS 1 ... Ontology processing apparatus, 2 ... Word vector preparation part, 3 ... Input / output part, 4 ... Ontology storage part, 5 ... Document storage part, 6 ... Information processing part, 7 ... Memory part, 11 ... Processing target selection part, 12 ... Coordinate conversion unit, 13 ... Distance calculation unit, 14 ... Ontology editing unit, 31 ... Input unit, 32 ... Output unit, 71 ... Ontology processing program.

Claims (3)

A plurality of words, a word vector holding unit that holds a plurality of word vectors corresponding to the plurality of words,
An ontology storage unit that stores the ontology,
From the ontology of the ontology storage unit, a seed word acquisition unit that acquires a word corresponding to a lower concept of a specified word as a seed word,
A coordinate conversion unit that extracts the word vector from the word vector holding unit and converts the extracted word vector into a conversion vector into a predetermined space in which cosine similarity can be approximated by a Euclidean distance;
A hyperplane determined from the result of the fitting process of the seed word, and a distance calculation unit that calculates a distance between the word of the conversion vector obtained by the conversion in the coordinate conversion unit,
Based on the distance calculated by the distance calculation unit, from the words held in the word vector holding unit, an additional candidate extraction unit that extracts a word that is a candidate to be additionally registered in the ontology of the ontology storage unit,
An ontology editing unit for adding a part or all of the words of the addition candidate extracted by the addition candidate extraction unit to the ontology in the ontology storage unit. The coordinate conversion unit calculates a seed center from an average of a plurality of word vectors of the seed word, and calculates a word vector corresponding to a word whose cosine similarity with the seed center obtained by the calculation is within a predetermined range. Normalizing each of the word vectors extracted from the word vector holding unit and each of the word vectors extracted from the word vector holding unit, such that the normalized seed center is the origin in the predetermined space; The ontology processing apparatus according to claim 1, wherein the ontology processing device converts the rotation vector around the origin and converts the rotation vector into the conversion vector. A plurality of words, a word vector holding unit that holds a plurality of word vectors corresponding to the plurality of words, and a computer including an ontology storage unit that stores an ontology,
From the ontology in the ontology storage unit, acquiring a word corresponding to a lower concept of a specified word as a seed word;
Extracting the word vector from the word vector holding unit, and converting the extracted word vector into a conversion vector into a predetermined space in which cosine similarity can be approximated by a Euclidean distance;
Calculating the distance between the hyperplane determined from the result of the fitting process of the seed word and the word of the conversion vector obtained by the conversion,
Extracting words that are candidates to be additionally registered in the ontology in the ontology storage unit, from the words stored in the word vector storage unit, based on the distance obtained by the calculation;
Adding, to the ontology in the ontology storage unit, a part or all of the extracted additional candidate words.