JP6495212B2 - Classifier learning device, table type classification device, method, and program - Google Patents

Description

  The present invention relates to a classifier learning device, table type classification device, method, and program for classifying table types of tabular data.

  With the development of computer technology, a large amount of tabular data described in HTML on the Web and tabular data on spreadsheets created by spreadsheet software etc. have become available. There are several types of tabular data, such as vertical or horizontal list-type tables, vertical or horizontal attribute-type tables, vertical or horizontal enumeration-type tables, matrix-type tables, and other layout tables. If you can understand this table type correctly, you can acquire knowledge applicable to a wide range of services such as information search and question answering.

Crestan, Eric and Patrick Pantel (2011). “Web-scale Table Census and Classification”. In: Proceedings of the 4th ACM International Conference on Web Search and Data Mining. ACM, pp. 545-554.

  Non-Patent Document 1 has been proposed as a conventional method for classifying table types.

  These conventional methods use features such as text length, text type (number, date, amount of money, etc.), text pattern (including specific symbols), etc., for text information described in cells in a table. However, the order of appearance of words and strings in the text described in the cells, the order of appearance of the cells in the rows and columns, the order of appearance of the rows, and the order of appearance of the columns are not considered.

  For example, the list type table is a table in which the attribute is arranged in the header portion and the attribute value is arranged in the data portion. In the example shown in FIG. 6 (A), when looking at the cell sequence from left to right for the second row, it is possible that the texts related to all the money amounts such as "monthly cost", "1800 yen" and "2000 yen" are arranged This table is characterized in that it is a horizontal list type. In addition, the fact that the similar columns are arranged in the second to fourth columns is a feature of the horizontal list type table.

  In the matrix type table, entities (line 回 線, option A, etc.) are described in the header part of rows and columns. In the example shown in FIG. 6 (B), looking at the cell sequence from left to right for the second row, it is a sequence of cells of “line」 ”and“ 1800 yen ”, and the semantic sequence is entity → attribute It is the feature of the matrix type table from the value and the continuation.

  The prior art can not distinguish between list-type and matrix-type tables having similar structures and text lengths as shown in FIGS. 6 (A) and 6 (B).

  The present invention has been made in view of the above circumstances, and takes into consideration the order of cells in a row, the order of rows in a sequence table in a column, and the order of columns in a table from the entire input tabular data. It is an object of the present invention to provide a classifier learning device, a table type classification device, a method, and a program capable of classifying table types of tabular data with high accuracy.

  In order to achieve the above object, a classifier learning device according to the present invention includes the training data set which is a set of a set of tabular data described as a matrix of cells and a correct answer label representing a type of the table. A cell encoding unit encoding, for each of the tabular data, for each cell of the table, a sequence of tokens included in the text information described in the cell as an input, into the continuous value vector; and the training data For each of the tabular data included in the set, for each row of the table, the sequence of continuous value vectors of each cell included in the row output by the cell encoding unit is input, and the row is converted to a continuous value vector For each row of the row encoding unit to be encoded and the tabular data included in the training data set, the continuous value of each cell included in the column output by the cell encoding unit for each column of the table Be A row coding unit encoding a sequence into a continuous value vector using a sequence of toru as input, and the row coding unit and the column coding unit outputting each of the tabular data included in the training data set A table encoding unit that inputs a sequence of continuous value vectors of each row and each column into a continuous value vector, and the table encoding unit performs coding on each of the tabular data included in the training data set; And a learning unit configured to learn a classifier for classifying a table type of tabular data based on the output continuous value vector and the correct answer label included in the training data set. .

  In the classifier learning method according to the present invention, the table format included in the training data set in which the cell encoding unit is a set of a set of tabular data described as a matrix of cells and a correct answer label representing a type of table. For each of the data, for each cell of the table, the sequence of tokens included in the text information described in the cell is input, and the cell is encoded into a continuous value vector, and the row coding unit generates the training data set. For each of the tabular data contained, for each row of the table, the sequence of continuous value vectors of each cell included in the row output by the cell encoding unit is input, and the row is encoded into the continuous value vector Of the continuous value vector of each cell included in the column, which the column encoding unit outputs, for each column of the table, for each of the tabular data included in the training data set; series As input, the columns are encoded into continuous value vectors, and the table encoding unit outputs, for each of the tabular data included in the training data set, each row and each row output by the row encoding unit and the column encoding unit. A table is encoded into a continuous value vector using a sequence of continuous value vectors of columns as input, and the learning unit outputs the continuous value vector output by the table encoding unit for each of the tabular data included in the training data set And learning a classifier for classifying a table type of tabular data based on the correct answer label included in the training data set.

  The table type classification apparatus according to the present invention continuously transmits cells for tabular data described as a matrix of cells, for each cell of the table, using a sequence of tokens included in the text information described in the cells as an input. A row is input using a cell encoding unit for encoding into a value vector, and the continuous value vector of each cell included in the row output from the cell encoding unit for each row of the table for the tabular data. A row encoding unit encoding a continuous value vector, and the continuous value vector of each cell included in the column output from the cell encoding unit for each column of the table for the tabular data, as an input; A column encoding unit that encodes a column into a continuous value vector, and a table of continuous value vectors of each row and each column output from the row encoding unit and the column encoding unit for the tabular data, as a table The continuous The tabular form based on a table encoding unit for encoding into a vector, a continuous value vector output from the table encoding unit, and a previously learned classifier for classifying the type of table of tabular data And a table type classification unit that classifies the type of data table.

  In the table type classification method according to the present invention, for the tabular data described by the cell encoding unit as a matrix of cells, for each cell of the table, a sequence of tokens included in the text information described in the cell The cells are encoded as continuous value vectors as input, and the row coding unit outputs the continuous value vector of each cell included in the row, which the cell coding unit outputs for each row of the table for the tabular data. Encoding the rows into a continuous value vector, and the column coding unit outputs, for each column of the table, the table format data, for each column of the table, the continuity of each cell included in the column The value vector is input and the column is encoded into a continuous value vector, and the table encoding unit outputs continuous value vectors of each row and each column output by the row encoding unit and the column encoding unit for the tabular data. If you input a series , The table is encoded into a continuous value vector, and the table type classification unit comprises: the continuous value vector output from the table coding unit; and a previously learned classifier for classifying the type of the tabular data table. Based on the table type of the tabular data is classified.

  Further, a program of the present invention is a program for causing a computer to function as the above-described classifier learning device or each unit constituting a table type classification device.

  As described above, according to the classifier learning apparatus, method, and program of the present invention, cells are encoded into continuous value vectors, and a sequence of continuous value vectors of each cell is input and a row is encoded into continuous value vectors. The column is encoded as a continuous value vector with the series of continuous value vectors of each cell as input, and the table is encoded into the continuous value vector with the series of continuous value vectors of each row and each column as input. By learning the classifier for classifying table types, the order of the cells in the row, the order of the rows in the order table in the columns, and the order of the columns in the table are considered from the entire input tabular data. Thus, it is possible to learn a classifier capable of accurately classifying table types of tabular data.

  Further, according to the table type classification device, method and program of the present invention, cells are encoded into continuous value vectors, a sequence of continuous value vectors of each cell is input, and rows are encoded into continuous value vectors. A sequence of continuous value vectors of is input to encode a column into a continuous value vector, and a sequence of continuous value vectors of each row and each column is input to encode a table into a continuous value vector, and the table type of tabular data is By classifying, the table type of the tabular data is accurately taken into consideration from the whole input tabular data, the order of the cells in the row, the order of the rows in the order table in the columns, and the order of the columns in the table. It can be classified.

It is a block diagram showing functional composition of a classifier learning device concerning an embodiment of the present invention. It is a figure for demonstrating the method to produce | generate a table coding vector. It is a block diagram showing functional composition of a table type classification device concerning an embodiment of the present invention. It is a flowchart figure of the classifier learning processing routine in the classifier learning device according to the embodiment of the present invention. It is a flowchart figure of the table type classification processing routine in the table type classification device concerning an embodiment of the present invention. (A) A figure showing an example of tabular form data of horizontal list type, A figure showing an example of tabular form data of matrix type (B).

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

<Configuration of Classifier Learning Device According to Embodiment of the Present Invention>
Next, the configuration of the classifier learning device according to the embodiment of the present invention will be described. As shown in FIG. 1, the classifier learning device 100 according to the present embodiment includes a CPU, a RAM, and a ROM storing a program for executing a classifier learning process routine described later and various data. It can be configured by a computer. The classifier learning device 100 functionally includes an input unit 10, an operation unit 20, and an output unit 40 as shown in FIG.

  The input unit 10 receives a training data set which is a set of a set of tabular data and a correct answer label representing a table type.

  Tabular data is a set of cells in matrix form, and each cell is given plain text information not containing HTML or text information containing HTML tags.

  The table types are, for example, vertical or horizontal list-type tables, vertical or horizontal attribute-type tables, vertical or horizontal enumeration-type tables, matrix-type tables, and other layout tables.

  As shown in FIG. 1, the operation unit 20 includes a cell encoding unit 22, a row encoding unit 24, a column encoding unit 26, a table encoding unit 28, and a learning unit 30.

  For each tabular data included in the training data set, the cell encoding unit 22 inputs, for each cell of the table, a sequence of tokens included in the text information described in the cell, and converts the cells into a continuous value vector Encode.

Specifically, for each table in the training data set, the cell encoding unit 22 performs the following processing for all cells c _{i, j} (row number i, column number j) included in the table, The coding vector h _{i, j} of each cell is output.

First, the cell encoding unit 22 divides the text information described in the cell c _{i, j} into a series of words (called tokens) by morphological analysis. For example, when "monthly cost" is received as plain text, it is divided into token series "monthly amount" and "cost". If you receive "<th><b> monthly </ b> cost </ th>" as HTML text, divide it into HTML tags and morphologically analyze the text in the tags <th> Divide into <b>, "Monthly", "</ b>", "Expense", and "</ th>". The text information may be divided for each character and one character may be used as a token.

Next, the cell coding unit 22 converts the sequence of tokens included in the cell c _{i, j} into a continuous value coded vector h _{i, j} .

  For encoding sequence data, LSTM shown in Non-Patent Document 2 or GRU shown in Non-Patent Document 3 can be used.

[Non-patent document 2]: S Hochreiter, J Schmidhuber. Long short-term memory. Neural computation 9 (8), 1735-1780, 1997.

[Non-patent document 3]: K. Cho, B. van Merrienboer, D. Bahdanau, and Y. Bengio. On the properties of neural machine translation: Encoder-decoder approaches. ArXiv preprint arXiv: 1409.1259, 2014.

  When GRU is used, h is updated by the following calculation each time token x is given, where each token of the token sequence is x and the encoding vector of Nc dimension is h.

Where σ is a sigmoid function,

Means the product of each element of the matrix. W _e , W _r , U _r , W _z , U _z , W, and U are parameter matrices common to cells. Assuming that the type of token is K types and the embedding dimension is E, x is a K-dimensional 1-of-K vector (only the portion corresponding to the token is 1, the other is 0). W _e is an E × K dimensional parameter matrix. The other W. is a parameter matrix of Nc × E dimensions. U. is a parameter matrix of Nc × Nc dimensions. The initial value of h is an Nc-dimensional zero vector.

The coding vector hi _{, j} may be output using LSTM instead of GRU in equation (2). Also, the token sequence is reversed, and the coding vector h _{i, j} ^b is similarly output, and the vector [hi _{, j} ; h _{i, j} ^b ] concatenated with the normal order coding vector is output to the cell coding unit 22. It is good also as an output. In this case, the size of the coding vector is 2Nc.

  The row encoding unit 24 inputs, for each of the tabular data included in the training data set, the series of continuous value vectors of each cell included in the row, which is output by the cell encoding unit 22 for each row of the table. Encode the rows into a continuous value vector.

Specifically, the row encoding unit 24 performs the following processing for all the rows (row number i) included in the table, and the series (j of the encoding vectors h _{i, j} of the respective cells included in the row) Convert = 1,2, ...) into a continuous-value encoding vector h _i . For example, h _i is initialized as an Nr-dimensional zero vector, and h _i is updated by substituting h _{i, j} as e in the above equation (2) each time the cell encoding vector h _{i, j} is given. . W _e , W _r , U _r , W _z , U _z , W, and U are common parameter matrices between rows, and W. is a parameter matrix of Nr × Nc dimensions. U. is a parameter matrix of Nr × Nr dimensions.

The coding vector may be output using LSTM instead of GRU in equation (2). In addition, the sequence of cells in the row is reversed, and the encoding vector h _i ^b is similarly output, and the vector [h _i ; h _i ^b ] concatenated with the encoding vector in the order is used as the output of the row encoding unit 24. Also good. In this case, the size of the coding vector is 2Nr.

  The column encoding unit 26 inputs, for each column of the table, the series of continuous value vectors of cells included in the column, which are output from the cell encoding unit 22 for each column of the table, for each of the tabular data included in the training data set Encode the columns into a continuous value vector as

Specifically, the column encoding unit 26 performs the following processing for all the columns (column number i) included in the table, and generates a series of coded vectors h _{i, j} of each cell included in the columns ( Convert i = 1, 2,...) into continuous value coded vector h _j . For example, h _j is initialized as a No-dimensional zero vector, and h is substituted for e = h _{i, j} and h = h _j in equation (2) each time the cell encoding vector h _{i, j} is given. Update _j . W _e , W _r , U _r , W _z , U _z , W, and U are parameter matrices common to the columns, and W. is a parameter matrix of No × Nc dimension. U. is a parameter matrix of No.times.No dimension.

The coding vector may be output using LSTM instead of GRU in equation (2). In addition, the sequence of cells in the column is reversed, and the encoding vector h _j ^b is similarly output, and the vector [h _j ; h _j ^b ] concatenated with the regular order encoding vector is output from the column encoding unit 26 As well. In this case, the size of the coding vector is 2No.

  The table encoding unit 28 receives a table of series of continuous value vectors of each row and each column output from the row encoding unit 24 and the column encoding unit 26 for each of the tabular data included in the training data set. Encode into a continuous value vector.

Specifically, first, the table encoding unit 28 converts a series (i = 1, 2,...) Of the encoding vector h _{i in} each _row into an encoding vector h _row of continuous values. For example, h _row is initialized as a No-dimension zero vector, and is updated by substituting e = h _i and h = h _row in equation (2) each time a row coding vector h _i is given. _{W e, W r, U r} , W z, U z, W, and U is a parameter matrix, W. Becomes Nt × Nr dimensional parameter matrix. U. is a parameter matrix of Nt × Nt dimensions.

The coding vector may be output using LSTM instead of GRU in equation (2). Alternatively, the sequence of rows may be reversed, the encoding vector h _row ^b may be similarly output, and the vector [h _row ; h _row ^b ] concatenated with the orderly encoding vector may be output. In this case, the size of the coding vector is 2Nt.

Then, the table encoding unit 28 converts the series (j = 1, 2,...) Of the encoding vector h _j of each column into the encoding vector h _col of continuous values. For example, h _col is initialized as an Nt-dimensional zero vector, and is updated by substituting e = h _i and h = h _col of equation (2) each time a column coding vector h _j is given. W and _{e, W r, U r,} W z, U z, W, U parameter matrix and (a different parameter from the conversion to coding vectors h _row of continuous values), W. The Nt × No dimension It becomes a parameter matrix of U. is a parameter matrix of Nt × Nt dimensions.

The coding vector may be output using LSTM instead of GRU in equation (2). In addition, the sequence of the rows may be reversed, and the encoding vector h _col ^b may be output in the same manner, and the vector [h _col ; h _col ^b ] concatenated with the encoding vector in normal order may be output. In this case, the size of the coding vector is 2Nt.

The table encoding unit 28 outputs [h _row ; h _col ] obtained by concatenating the vectors output above as a table encoding vector h.

  The learning unit 30 determines the type of tabular data table based on the continuous value vector output by the tabular coding unit 28 for each of the tabular data included in the training data set and the correct answer label included in the training data set. Learn a classifier to classify

  Specifically, learning unit 30 receives, for each of the tables included in the training data set, table coding vector h (dimension number 2Nt) output from table coding unit 28 according to the following equation (3) , Class classification of table types (number of classes = C).

Here, W _cls is a parameter matrix of size 2Nt × C, softmax is a softmax function, and y is a vector of dimension C. In this embodiment, C = 6, and there are six types of classes: horizontal list type, vertical list type, horizontal attribute type, vertical attribute type, matrix type, and others.

  Then, for each of the tables included in the training data set, the learning unit 30 generates a list in the horizontal direction, a list in the vertical direction, and a horizontal attribute in the correct class t (t = 1,... C) of the table. The loss L is calculated by the cross entropy function of equation (4) using the vertical attribute type (meaning matrix type, etc.) and the output y of the class classification for the table.

  Then, the learning unit 30 calculates cell loss 22 for each table obtained by the above equation (4) using the stochastic gradient descent method, cell coding unit 22, row coding unit 24, column coding unit 26, table coding unit 28. Optimization is performed for each parameter matrix in the row direction and the column direction. Note that the present invention does not depend on the optimization method, and other optimization methods for other neural network parameters may be used.

FIG. 2 shows an example of cell / row / table coding vector. In this example, the cell encoding unit 22 extracts a token sequence of <td>, 1000, </ td> from the HTML text of “<td> 1000 </ td>” described in the cell c ₃ , ₁ by the GRU. _Encode into cell vector h _3,1 . Next, the row encoding unit 24 encodes the arrangement of cells in each row into a row vector. In this example, a series of cell vectors in the third row is encoded by GRU, acquiring row vector h _3. Then, the table encoding unit 28 encodes the series of rows by GRU to obtain a table vector h _row . Similarly, h _col is obtained from the series of columns, and classification is performed using a connected vector of [h _row ; h _col ].

<Configuration of Table Type Classification Device According to Embodiment of the Present Invention>
Next, the configuration of the table type classification device according to the embodiment of the present invention will be described. As shown in FIG. 3, the table type classification device 150 according to the present embodiment includes a CPU, a RAM, and a ROM storing a program for executing a table type classification processing routine described later and various data. It can be configured by a computer. The table type classification device 150 functionally includes an input unit 60, an operation unit 70, and an output unit 90 as shown in FIG.

  The input unit 60 receives tabular data to be classified.

  As shown in FIG. 3, the arithmetic unit 70 includes a cell encoding unit 72, a row encoding unit 74, a column encoding unit 76, a table encoding unit 78, and a table type classification unit 80.

For the input tabular data, for each cell of the table, the sequence of tokens included in the text information described in the cell is input as in the cell encoding unit 22 for the input tabular data, The cells are encoded into a continuous value vector h _{i, j} . Here, the parameter matrix used in the conversion of the continuous value into the coding vector h _{i, j} is the one optimized by the classifier learning device 100.

For the input tabular data, the row encoding unit 74 outputs the continuous value vector of the cells included in the row, which is output from the cell encoding unit 72 for each row of the table, similarly to the row encoding unit 24. With the sequence as input, the rows are encoded into a continuous value vector h _i . Here, the parameter matrix used in the conversion to the coding vector h _i of each row is one optimized by the classifier learning device 100.

The column encoding unit 76, for the input tabular data, as in the column encoding unit 26, for each column of the table, the continuous value vector of the cells included in the column, which the cell encoding unit 72 outputs. The sequence is encoded into a continuous value vector h _j using as input. Here, the parameter matrix used in the conversion to the encoding vector h _j of each column is optimized by the classifier learning device 100.

The table encoding unit 78 inputs a series of continuous value vectors of each row and each column output from the row encoding unit 74 and the column encoding unit 76 for the input tabular data in the same manner as the table encoding unit 28. Encode the table into a continuous value vector h. Here, the parameter matrix used in the conversion of the continuous value into the coding vector h _row and the conversion of the continuous value into the coding vector h _row is the one optimized by the classifier learning device 100.

  The table type classification unit 80 receives the table coding vector h (dimension number 2Nt) output from the table coding unit 78 as input for the input tabular data, in the same manner as the learning unit 30 according to the above equation (3). , Class classification of table types (number of classes = C).

<Operation of Classifier Learning Device According to Embodiment of the Present Invention>
Next, the operation of the classifier learning device 100 according to the embodiment of the present invention will be described. In the classifier learning device 100, when the training data set is received by the input unit 10, the classifier learning processing routine shown in FIG. 4 is executed by the classifier learning device 100.

  First, in step S100, for each of the tabular data included in the input training data set, for each cell of the table, the sequence of cells included in the text information described in the cell is input, and the cells are continued Encode into a value vector.

  In step S102, for each of the tabular data included in the input training data set, for each row of the table, the sequence of continuous value vectors of cells included in the row obtained in step S100 is input. Encode the rows into a continuous value vector.

  Also, for each of the tabular data included in the training data set, for each column of the table, the series of continuous value vectors of the cells included in the column obtained in step S100 above are input, and the columns are continuous values Encode into a vector.

  In step S104, for each of the tabular data included in the training data set, the sequence of the continuous value vector of each row and column obtained in step S102 is input to encode the table into a continuous value vector.

  In step S106, the table coding vector h (dimension number 2Nt) obtained in step S104 is input for each of the tabular data included in the training data set, and the table classification is classified according to the above equation (3). Perform (Number of classes = C).

  Then, in step S108, for each of the tables included in the training data set, using the correct class of the table and the output y of the class classification in step S106 for the table, the cross entropy function of the above equation (4) The loss L is calculated by Then, with regard to the loss L for each table, the parameter coding matrix in the row direction and the column direction of the cell coding unit 22, the row coding unit 24, the column coding unit 26, and the table coding unit 28 is optimized by the stochastic gradient descent method. And output from the output unit 40, and the classifier learning processing routine is ended.

<Operation of Table Type Classification Device According to Embodiment of the Present Invention>
Next, the operation of the table type classification device 150 according to the embodiment of the present invention will be described. When the table type classification device 150 receives tabular data by the input unit 60, the table type classification device 150 executes the table type classification processing routine shown in FIG.

  First, in step S150, for input tabular data, for each cell of the table, the sequence of tokens included in the text information described in the cell is input, and the cell is encoded into a continuous value vector.

  In step S152, for the input tabular data, for each row of the table, the sequence of continuous value vectors of cells included in the row obtained in step S150 is input to encode the row into a continuous value vector Do.

  Also, for the input tabular data, for each column of the table, the series is encoded with the series of continuous value vectors of the cells included in the column obtained in step S150 above into the continuous value vector .

  In step S154, the table is encoded into a continuous value vector, with the input tabular data being input, the series of continuous value vectors of each row and each column obtained in step S152.

  In step S156, the table coding vector h (dimension number 2Nt) obtained in step S154 is input to the input tabular data, and the table classification class classification (class number = C) according to the above equation (3) And the output unit 90 outputs the table type classification processing routine.

  As described above, according to the table type classification device according to the embodiment of the present invention, a cell is encoded into a continuous value vector, a sequence of the continuous value vector of each cell is input, and a row is coded into the continuous value vector. The column is encoded as a continuous value vector with the series of continuous value vectors of each cell as input, and the table is encoded into the continuous value vector with the series of continuous value vectors of each row and each column as input. By classifying the type of table, the table type data table is considered taking into consideration the order of the cells in the row, the order of the rows in the order table in the columns, and the order of the columns in the table from the entire input table data. The types can be classified with high accuracy.

  Also, according to the classifier learning device according to the embodiment of the present invention, cells are encoded into continuous value vectors, and a sequence of continuous value vectors of each cell is input, and rows are encoded into continuous value vectors, A sequence of continuous value vectors of is input to encode a column into a continuous value vector, and a sequence of continuous value vectors of each row and each column is input to encode a table into a continuous value vector, and the table type of tabular data is By learning the classifier for classification, the tabular form taking into consideration the order of the cells in the row, the order of the rows in the order table in the columns, and the order of the columns in the table from the entire input tabular data It is possible to learn a classifier capable of accurately classifying table types of data.

  The present invention is not limited to the above-described embodiment, and various modifications and applications can be made without departing from the scope of the present invention.

  For example, in the above embodiment, the classifier learning device and the table type classification device are separately provided, but the classifier learning device and the table type classification device can be realized by one device. You may

  In the present specification, although the program has been described as an embodiment in which the program is pre-installed, the program can be provided by being stored in a computer readable recording medium, or provided via a network. Is also possible.

  The present invention can be used for knowledge acquisition and information retrieval from tabular data whose precision can be improved by understanding table types.

10, 60 input unit 20, 70 operation unit 22, 72 cell encoding unit 24, 74 row encoding unit 26, 76 column encoding unit 28, 78 table encoding unit 30 learning unit 40, 90 output unit 80 table type classification Part 100 Classifier learning device 150 Table type classification device

Claims (5)

For each cell in the table, for each cell in the table, for each cell in the table, for each cell in the table, for each of the tabular data included in the training data set that is a set of sets of tabular data described as a matrix of cells and correct A cell encoding unit that encodes a cell into a continuous value vector with a sequence of tokens included in the processed text information as an input;
For each of the tabular data included in the training data set, for each row of the table, the row is continuously input using the series of continuous value vectors of each cell included in the row, which is output by the cell encoding unit. A row encoding unit encoding the value vector;
For each of the tabular data included in the training data set, for each column of the table, the column is input using the sequence of continuous value vectors of each cell included in the column, which is output by the cell encoding unit. A column encoding unit that encodes a continuous value vector;
For each of the tabular data included in the training data set, a sequence of continuous value vectors of each row and each row output by the row coding unit and the column coding unit is input, and a table is coded as a continuous value vector A table encoding unit to convert
For each of the tabular data included in the training data set, the table type of the tabular data is determined based on the continuous value vector output by the table encoding unit and the correctness label included in the training data set. A learning unit that learns a classifier for classification;
A classifier learning device characterized by having: A cell encoding unit encoding, into tabular values represented as a matrix of cells, for each cell of the table, a sequence of tokens included in text information described in the cell as an input and encoding the cells into a continuous value vector When,
A row coding unit encoding the row into a continuous value vector by using the continuous value vector of each cell included in the row output from the cell coding unit for each row of the table for the tabular data; ,
A column coding unit which encodes the columns into continuous value vectors by using the continuous value vector of each cell included in the column output from the cell encoding unit for each column of the table for the tabular data When,
A table encoding unit that encodes a table into a continuous value vector, with the tabular data, the series of continuous value vectors of each row and each column output by the row coding unit and the column coding unit being input;
A table type classification that classifies the type of the tabular data table based on the continuous value vector outputted by the table encoding unit and a previously learned classifier for classifying the type of the tabular data table. Department,
A table type classification device characterized by having. For each cell of the table, for each of the tabular data included in the training data set, the cell encoding unit is a set of a set of tabular data described as a matrix of cells and a correct answer label representing the type of the table. And the cell is encoded into a continuous value vector with the sequence of tokens included in the text information described in the cell as an input,
A row coding unit outputs, for each row of the table, for each of the tabular data included in the training data set, a series of continuous value vectors of each cell included in the row, which is output by the cell coding unit. Encode the rows into a continuous value vector as input
A sequence of continuous value vectors of each cell included in the column, which the column encoding unit outputs, for each column of the table, for each of the tabular data included in the training data set. Encode the columns into a continuous value vector, taking
A table encoding unit is input for each of the tabular data included in the training data set, using a sequence of continuous value vectors of each row and each column output by the row encoding unit and the column encoding unit as a table. Encode into a continuous value vector,
The learning unit generates the tabular data based on the continuous value vector output from the tabular coding unit for each of the tabular data included in the training data set, and the correctness label included in the training data set. Train a classifier to classify table types,
Classifier learning method. For tabular data described as a matrix of cells, the cell encoding unit inputs, for each cell of the table, a sequence of tokens included in the text information described in the cell and codes the cells into a continuous value vector Turn
The row coding unit encodes the row into a continuous value vector by using the continuous value vector of each cell included in the row output from the cell coding unit for each row of the table for the tabular data ,
A column encoding unit is a unit for encoding a column as a continuous value vector, with the continuous value vector of each cell included in the column output by the cell encoding unit for each column of the table for the tabular data Turn
The table encoding unit encodes a table into a continuous value vector, with the tabular data, the sequence of continuous value vectors of each row and each column output by the row coding unit and the column coding unit being input,
A table type classification unit generates the table type of the tabular data based on the continuous value vector output by the table encoding unit and a previously learned classifier for classifying the type of the tabular data table. Table type classification method.   A program for causing a computer to function as each part of the classifier learning device according to claim 1 or the table type classification device according to claim 2.