Document dataset analysis device and document dataset analysis method

Abstract

To analyze features common to two document datasets, features peculiar to one document dataset, and features peculiar to the other document dataset by taking into account time information included in each piece of document data.SOLUTION: A three-dimensional non-negative tensor T1 generated from a document dataset D1 is factorized into two-dimensional matrices U1, V1, W1, and a three-dimensional non-negative tensor T2 generated from a document dataset D2 is factorized into two-dimensional matrices U1, V1, W1. The two-dimensional matrix U1 is composed of a submatrix U1c corresponding to a topic common to the document datasets D1, D2, and a submatrix U1d corresponding to a topic peculiar to the document dataset D1. A two-dimensional matrix U2 is composed of a submatrix U2c corresponding to a topic common to the document datasets D1, D2, and a submatrix U2d corresponding to a topic peculiar to the document dataset D2.SELECTED DRAWING: Figure 3

Description

Technical Field

[0001] The present invention relates to a document data set analysis device for analyzing a document data set, a document data set analysis method, and a document data set analysis program.

Background Art

[0002] As a technique for extracting topics from document data and grasping the content of the document data, the LDA (Latent Dirichlet Allocation) method based on a probabilistic model is known (see Non-Patent Document 1).

Prior Art Documents

Non-Patent Documents

[0003]

Non-Patent Document 1

Non-Patent Document 2

Summary of the Invention

[0004] When comparing two document datasets, it is important to understand the common features of both datasets, the features unique to one dataset, and the features unique to the other dataset. In particular, if each document contains temporal information, it is also important to understand how these three features change over time.

[0005] One example of such a document dataset is a dataset consisting of patent documents that contain time information such as filing dates and registration dates. When comparing a document dataset consisting of patent documents from company A with one consisting of patent documents from company B, the temporal changes in these three characteristics provide valuable information for understanding the evolution of the technological trends of both companies. However, the LDA method described in Non-Patent Document 1 cannot meet this need. The same can be said for academic papers that contain time information such as acceptance dates and publication dates, and news articles that contain time information such as publication dates and release dates.

[0006] Joint Non-negative Matrix Factorization (JMF) is a known method for comparing two document datasets (see Non-Patent Literature 2). However, JMF cannot perform analysis that takes into account the temporal information contained in the document data. Therefore, JMF also cannot adequately meet the above needs.

[0007] One aspect of the present invention has been made in view of the above-mentioned problems, and its objective is to realize a document dataset analysis technique that can analyze features common to two document datasets, features unique to one document dataset, and features unique to the other document dataset, taking into account the temporal information contained in each document dataset. [Means for solving the problem]

[0008] A document dataset analysis device according to embodiment 1 of the present invention comprises at least one processor. The processor performs a generation process and a decomposition process.

[0009] Here, the generation process involves generating a three-dimensional non-negative tensor X1 from a document dataset D1 consisting of document data containing time information, which has the dimensions of document data, word, and time, and whose elements are the frequency of occurrence of each word in each document data corresponding to each time. Simultaneously, it generates a three-dimensional non-negative tensor X2 from a document dataset D2 consisting of document data containing time information, which has the dimensions of document data, word, and time, and whose elements are the frequency of occurrence of each word in each document data corresponding to each time.

[0010] Furthermore, the decomposition process involves factorizing a three-dimensional non-negative tensor X1 into a two-dimensional matrix U1 with the dimensions of topic and word, where the elements are the frequency of occurrence of each word in each topic; a two-dimensional matrix V1 with the dimensions of document data and topic, where the elements are the weights of each topic in each document data; and a two-dimensional matrix W1 with the dimensions of time and topic, where the elements are the weights of each topic at each time; and also factorizing a three-dimensional non-negative tensor X2 into a two-dimensional matrix U2 with the dimensions of topic and word, where the elements are the frequency of occurrence of each word in each topic; a two-dimensional matrix V2 with the dimensions of document data and topic, where the elements are the weights of each topic in each document data; and a two-dimensional matrix W2 with the dimensions of time and topic, where the elements are the weights of each topic at each time.

[0011] Here, the two-dimensional matrix U1 is composed of a submatrix U1c corresponding to topics common to document datasets D1 and D2, and a submatrix U1d corresponding to topics specific to document dataset D1. The two-dimensional matrix U2 is composed of a submatrix U2c corresponding to topics common to document datasets D1 and D2, and a submatrix U2d corresponding to topics specific to document dataset D2.

[0012] With the above configuration, it is possible to analyze the features common to document datasets D1 and D2, the features unique to document dataset D1, and the features unique to document dataset D2, taking into account the temporal information contained in each document data. Specifically, for each topic common to document datasets D1 and D2, and each topic unique to document dataset D1, a matrix U1 showing the frequency of occurrence of each word in document dataset D1 can be obtained. Similarly, for each topic common to document datasets D1 and D2, and each topic unique to document dataset D2, a matrix U2 showing the frequency of occurrence of each word in document dataset D2 can be obtained. Furthermore, for each topic common to document datasets D1 and D2, and each topic unique to document dataset D1, a matrix V1 showing the weight of each document data in document dataset D1 can be obtained. Similarly, for each topic common to document datasets D1 and D2, and each topic unique to document dataset D2, a matrix V2 showing the weight of each document data in document dataset D2 can be obtained. Furthermore, for each topic common to document datasets D1 and D2, and each topic specific to document dataset D1, a matrix W1 can be obtained showing the time-dependent weights in document dataset D1. Similarly, for each topic common to document datasets D1 and D2, and each topic specific to document dataset D2, a matrix W2 can be obtained showing the time-dependent weights in document dataset D2.

[0013] In the document dataset analysis device according to aspect 2 of the present invention, in addition to the configuration of the document dataset analysis device according to aspect 1 of the present invention, the following configuration is adopted. That is, in the decomposition process, the processor determines the values ​​of the elements of the two-dimensional matrices U1, U2, V1, V2, W1, W2 in such a way that the value increases as the submatrix U1c and submatrix U2c become more similar, and decreases as the submatrix U1d and submatrix U2d become more similar.

[0014] According to the above configuration, the values ​​of the elements of the two-dimensional matrices U1, U2, V1, V2, W1, and W2 can be determined efficiently.

[0015] In the document dataset analysis device according to aspect 3 of the present invention, in addition to the configuration of the document dataset analysis device according to any one of aspects 1 to 2 of the present invention, the following configuration is adopted. That is, the processor further executes output processing which outputs a table T1c in which words are arranged in the order of occurrence frequency shown by submatrix U1c for each topic common to document datasets D1 and D2, a table T1d in which words are arranged in the order of occurrence frequency shown by submatrix U1d for each topic unique to document dataset D1, a table T2c in which words are arranged in the order of application frequency shown by submatrix U2c for each topic common to document datasets D1 and D2, and a table T2d in which words are arranged in the order of occurrence frequency shown by submatrix U2d for each topic unique to document dataset D2.

[0016] With the above configuration, the user can be informed of the frequency of occurrence of each word in document dataset D1 for each topic common to document datasets D1 and D2, and for each topic specific to document dataset D1. Similarly, the user can be informed of the frequency of occurrence of each word in document dataset D2 for each topic common to document datasets D1 and D2, and for each topic specific to document dataset D2.

[0017] In the document dataset analysis device according to aspect 4 of the present invention, in addition to the configuration of the document dataset analysis device according to any one of aspects 1 to 3 of the present invention, the following configuration is adopted. That is, the processor further performs output processing to output a graph G1 representing the time change of the weights shown by matrix W1 for each topic common to document datasets D1 and D2, and each topic specific to document dataset D1, and a graph G2 representing the time change of the weights shown by matrix W2 for each topic common to document datasets D1 and D2, and each topic specific to document dataset D2.

[0018] With the above configuration, the user can be informed of the time-dependent changes in the weights in document dataset D1 for each topic common to both document datasets D1 and D2, and for each topic specific to document dataset D1. Similarly, the user can be informed of the time-dependent changes in the weights in document dataset D2 for each topic common to both document datasets D1 and D2, and for each topic specific to document dataset D2.

[0019] A document dataset analysis method according to aspect 5 of the present invention includes a generation process performed by at least one processor and a decomposition process performed by the processor.

[0020] Here, the generation process involves generating a three-dimensional non-negative tensor X1 from a document dataset D1 consisting of document data containing time information, which has the dimensions of document data, word, and time, and whose elements are the frequency of occurrence of each word in each document data corresponding to each time. Simultaneously, it generates a three-dimensional non-negative tensor X2 from a document dataset D2 consisting of document data containing time information, which has the dimensions of document data, word, and time, and whose elements are the frequency of occurrence of each word in each document data corresponding to each time.

[0021] Also, the decomposition process factorizes the three-dimensional non-negative tensor X1 into a two-dimensional matrix having dimensions of topics and words, i.e., a two-dimensional matrix U1 with the element being the frequency of occurrence of each word in each topic, a two-dimensional matrix V1 having dimensions of document data and topics, i.e., a two-dimensional matrix with the element being the weight of each topic in each document data, and a two-dimensional matrix W1 having dimensions of time and topics, i.e., a two-dimensional matrix with the element being the weight of each topic at each time. At the same time, the three-dimensional non-negative tensor X2 is factorized into a two-dimensional matrix having dimensions of topics and words, i.e., a two-dimensional matrix U2 with the element being the frequency of occurrence of each word in each topic, a two-dimensional matrix V2 having dimensions of document data and topics, i.e., a two-dimensional matrix with the element being the weight of each topic in each document data, and a two-dimensional matrix W2 having dimensions of time and topics, i.e., a two-dimensional matrix with the element being the weight of each topic at each time.

[0022] Here, the two-dimensional matrix U1 is composed of a sub-matrix U1c corresponding to the topics common to the document data sets D1 and D2 and a sub-matrix U1d corresponding to the topics unique to the document data set D1. The two-dimensional matrix U2 is composed of a sub-matrix U2c corresponding to the topics common to the document data sets D1 and D2 and a sub-matrix U2d corresponding to the topics unique to the document data set D2.

[0023] According to the above configuration, similar to the document data set analysis device according to Embodiment 1 of the present invention, it is possible to analyze the features common to the document data sets D1 and D2, the features unique to the document data set D1, and the features unique to the document data set D2, taking into account the time information possessed by each document data, and thus there is an effect that such analysis can be performed.

[0024] The document data set analysis program according to Embodiment 6 of the present invention is a program for causing a computer to operate as the document data set analysis device according to any one of Embodiments 1 to 4, and causes the computer to execute each of the above processes (generation process, decomposition process, output process).

[0025] According to the above configuration, similar to the document data set analysis device according to Embodiment 1 of the present invention, it is possible to analyze the features common to the document data sets D1 and D2, the features unique to the document data set D1, and the features unique to the document data set D2, taking into account the time information possessed by each document data, and thus there is an effect.

Effect of the Invention

[0026] According to one aspect of the present invention, it is possible to analyze the features common to two document data sets, the features unique to one document data set, and the features unique to the other document data set, taking into account the time information possessed by each document data.

Brief Description of the Drawings

[0027] [Figure 1] It is a block diagram showing the configuration of a document data set analysis device according to an embodiment of the present invention. [Figure 2] It is a flowchart showing the flow of a document data set analysis method according to an embodiment of the present invention. [Figure 3] It is a diagram schematically showing the structure of various data generated by the document data set analysis method shown in FIG. 2. [Figure 4] It is a diagram showing an example of the present invention. (a) shows a table in which the top 10 words in the order of appearance frequency shown by the partial matrix U1c are arranged for each of the topics 1 to 9 common to both companies. (b) shows a table in which the top 10 words in the order of appearance frequency shown by the partial matrix U1d are arranged for the topics 10 to 18 unique to Company A. [Figure 5] It is a diagram showing an example of the present invention. (a) shows a table in which the top 10 words in the order of appearance frequency shown by the partial matrix U2c are arranged for each of the topics 1 to 9 common to both companies. (b) shows a table in which the top 10 words in the order of appearance frequency shown by the partial matrix U2d are arranged for the topics 10 to 18 unique to Company B. [Figure 6]This figure shows one embodiment of the present invention. (a) shows a stacked line graph showing the time change of the weights represented by matrix W1 for topics 1 to 9 common to both companies and topics 10 to 18 specific to company A. (b) shows a stacked line graph showing the time change of the weights represented by matrix W1 for topics 1 to 9 common to both companies and topics 10 to 18 specific to company B. [Modes for carrying out the invention]

[0028] (Configuration of the document dataset analysis system) The configuration of the document dataset analysis device 1 according to one embodiment of the present invention will be described with reference to Figure 1. Figure 1 is a block diagram showing the configuration of the document dataset analysis device 1.

[0029] The document dataset analysis device 1 is implemented using a general-purpose computer and, as shown in Figure 1, comprises a processor 11, a primary memory 12, a secondary memory 13, an input / output interface 14, and a bus 15. The processor 11, primary memory 12, secondary memory 13, and input / output interface 14 are interconnected via the bus 15.

[0030] The secondary memory 13 stores the document dataset analysis program P1 and document datasets D1 and D2. The document dataset analysis program P1 is a program that causes the computer to execute the document dataset analysis method S1, which will be described later. Document datasets D1 and D2 are sets of multiple document data each containing time information. The time information contained in the document data can be any information that represents the time corresponding to that document data, for example, it may be information that represents the time the document data was created, or it may be time information that represents the time the document data was published. Furthermore, the precision of the time information is arbitrary and may be yearly precision, monthly precision, day precision, or even higher precision (hourly precision, minute precision, second precision, etc.).

[0031] The processor 11 loads the document dataset analysis program P1 stored in the secondary memory 13 onto the primary memory 12. Then, the processor 11 executes each step of the document dataset analysis method S1, which will be described later, according to the instructions contained in the document dataset analysis program P1 loaded onto the primary memory 12. During this process, the processor 11 refers to the document datasets D1 and D2 stored in the secondary memory 13.

[0032] Examples of devices that can be used as processor 11 include CPUs (Central Processing Units) and GPUs (Graphical Processing Units). Examples of devices that can be used as primary memory 12 include semiconductor RAM (Random Access Memory). Examples of devices that can be used as secondary memory 13 include HDDs (Hard Disk Drives).

[0033] Input and / or output devices are connected to the input / output interface 14. An example of an input device connected to the input / output interface 14 is a keyboard. An example of an output device connected to the input / output interface 14 is a display. The display is used to output various screens obtained by executing the document dataset analysis method S1, which will be described later.

[0034] Examples of interfaces that can be used as input / output interface 14 include PCI (Peripheral Component Interconnect) interfaces and USB (Universal Serial Bus) interfaces.

[0035] The document dataset analysis program P1 can be recorded on a computer-readable recording medium, such as a tangible, non-temporary recording medium. Examples of such recording media include, in addition to the secondary memory 13 mentioned above, tapes, disks, cards, semiconductor memory, and programmable logic circuits.

[0036] (Flowchart of document dataset analysis method) The flow of the document dataset analysis method S1 according to one embodiment of the present invention will be described with reference to Figures 2 and 3. Figure 2 is a flowchart showing the flow of the document dataset analysis method S1. Figure 3 is a schematic diagram showing the structure of various data generated by the document dataset analysis method S1.

[0037] In the following explanation, the r-th word out of the R words under consideration will be referred to as "word r". Similarly, the s-th document data out of the S1 document data in document dataset D1 will be referred to as "document data s∈D1". Furthermore, the s-th document data out of the S2 document data in document dataset D2 will be referred to as "document data s∈D2". Finally, the t-th time out of the T times under consideration will be referred to as "time t".

[0038] As shown in Figure 2, the document dataset analysis method S1 includes a generation process S11, a decomposition process S12, a first output process S13, and a second output process S14. The generation process S11, the decomposition process S12, the first output process S13, and the second output process S14 are each executed by the processor 11 according to instructions contained in the document dataset analysis program P1 expanded on the primary memory 12.

[0039] The generation process S11 generates a three-dimensional non-negative tensor X1 from document dataset D1 and a three-dimensional non-negative tensor X2 from document dataset D2. Here, the three-dimensional non-negative tensor X1 is a three-dimensional non-negative tensor having dimensions of word, document data, and time. The elements X1(r,s,t) of the three-dimensional non-negative tensor X1 are given by the frequency of occurrence of word r in document data s∈D1 corresponding to time t (e.g., tf-idf value). Note that if document data s corresponding to time t does not exist in document dataset D1, the element X1(r,s,t) is set to 0. Similarly, the elements X2(r,s,t) of the three-dimensional non-negative tensor X2 are given by the frequency of occurrence of word r in document data s∈D2 corresponding to time t (e.g., tf-idf value). Note that if document data s corresponding to time t does not exist in document dataset D2, the element X2(r,s,t) is set to 0.

[0040] The decomposition process S12 is a process that factorizes a three-dimensional non-negative tensor X1 into two-dimensional matrices U1, V1, and W1, and also factorizes a three-dimensional non-negative tensor X2 into two-dimensional matrices U2, V2, and W2. Here, the two-dimensional matrices U1 and U2 are two-dimensional matrices with R rows and K columns, respectively, having the dimensions of topic and word. The elements U1(r,k) of the two-dimensional matrix U1 and U2(r,k) of the two-dimensional matrix U2 represent the frequency of occurrence of word r in topic k, respectively. The two-dimensional matrix V1 is a two-dimensional matrix with K rows and S1 columns, having the dimensions of document data and topic. The element V1(s,k) of the two-dimensional matrix V1 represents the weight of topic k in document data s∈D1. The two-dimensional matrix V2 is a two-dimensional matrix with K rows and S2 columns, having the dimensions of document data and topic. The element V(s,k) of the two-dimensional matrix V2 represents the weight of topic k in the document data s∈D2. Furthermore, the two-dimensional matrices W1 and W2 are K x T two-dimensional matrices with time and topic dimensions, respectively. The elements W1(t,k) of the two-dimensional matrix W1 and W2(t,k) of the two-dimensional matrix W2 represent the weight of topic k at time t, respectively.

[0041] The two-dimensional matrix U1 is composed of a submatrix U1c with R rows and Kc columns corresponding to Kc topics common to both document dataset D1 and document dataset D2, and a submatrix U1d with R rows and Kd columns corresponding to Kd (= K-Kc) topics unique to document dataset D1. Similarly, the two-dimensional matrix U2 is composed of a submatrix U2c with R rows and Kc columns corresponding to Kc topics common to both document dataset D1 and document dataset D2, and a submatrix U2d with R rows and Kd columns corresponding to Kd topics unique to document dataset D2. Hereafter, the Kc topics common to both document dataset D1 and document dataset D2 will also be referred to as "common topics." Furthermore, the Kd topics unique to document dataset D1 will be referred to as "unique topics of document dataset D1," and the Kd topics unique to document dataset D2 will be referred to as "unique topics of document dataset D2."

[0042] In the decomposition process S12, the processor 11 determines the values ​​of the elements of the two-dimensional matrices U1, U2, V1, V2, W1, W2 so as to minimize a loss function L whose value decreases as the submatrices U1c and U2c become more similar, and whose value increases as the submatrices U1d and U2d become more similar. Such a loss function L can be given, for example, by the following equation.

number

[0043] Here, ||·||² represents the Euclidean norm, and ||·|| F represents the Frobenius norm, and ||·||1 represents the Manhattan distance. Furthermore, α is a hyperparameter for adjusting the influence of the similarity between submatrices U1c and U2c, and β is a hyperparameter for adjusting the influence of the similarity between submatrices U1d and U2d. Additionally, the element values ​​of the two-dimensional matrices U1, U2, V1, V2, W1, W2 that minimize the loss function L are determined to satisfy the following normalization conditions.

number

[0044] The first output process S13 is the process of outputting tables T1c, T1d, T2c, and T2d, corresponding to the submatrices U1c, U1d, U2c, and U2d, respectively, to a display or the like. Here, table T1c is a table in which words for each common topic of document datasets D1 and D2 are arranged in the order of occurrence shown by submatrice U1c. Table T1d is a table in which words for each specific topic of document dataset D1 are arranged in the order of occurrence shown by submatrice U1d. Table T2c is a table in which words for each common topic of document datasets D1 and D2 are arranged in the order of occurrence shown by submatrice U2c. Table T2d is a table in which words for each specific topic of document dataset D1 are arranged in the order of occurrence shown by submatrice U2d. Specific examples of tables T1c, T1d, T2c, and T2d output to a display or the like will be explained in the embodiment.

[0045] The second output process S14 is the process of outputting graphs G1 and G2, corresponding to matrices W1 and W2, respectively, to a display or the like. Here, graph G1 is a graph showing the time change of the weights indicated by matrix W1 for each common topic of document datasets D1 and D2, and each specific topic of document dataset D1. Graph G2 is a graph showing the time change of the weights indicated by matrix W2 for each common topic of document datasets D1 and D2, and each specific topic of document dataset D2. Specific examples of graphs G1 and G2 output to a display or the like will be explained in the embodiment.

[0046] The document dataset analysis method S1 may further include a third output process that outputs graphs H1 and H2, corresponding to matrices V1 and V2, respectively, to a display or the like. Here, graph H1 is a graph showing the weight distribution (distribution on document dataset D1) shown by matrix W1 for each common topic of document datasets D1 and D2, and each specific topic of document dataset D1. Graph H2 is a graph showing the weight distribution (distribution on document dataset D2) shown by matrix W2 for each common topic of document datasets D1 and D2, and each specific topic of document dataset D2.

[0047] (Examples) The inventors performed the document dataset analysis method S1 described above on document dataset D1, consisting of 1650 patent documents filed by Company A between 2001 and 2021, and document dataset D2, consisting of 1277 patent documents filed by Company B between 2001 and 2021. Here, the target words were compound words formed by the combination of two or more nouns, with a frequency of 5 to 1000 in the entire document. The target time period was the year of filing. The number of target time periods T was T=21. The number of common topics Kc and the number of unique topics Kd were set to Kc=Kd=10. The hyperparameters α and β included in the loss function L were set to α=β=1.

[0048] The tables T1c, T1d, T2c, and T2d output to the display by executing the document dataset analysis method S1 are shown in Figures 4 and 5.

[0049] Figure 4(a) shows table T1c, which lists the top 10 words in order of frequency of occurrence as shown by submatrix U1c for each of topics 1 to 9 common to both companies, and Figure 4(b) shows table T1d, which lists the top 10 words in order of frequency of occurrence as shown by submatrix U1d for topics 10 to 18 specific to company A. By referring to these tables, users can understand the keywords that indicate the trends in company A's technology, separating them into topics common to company B and topics specific to company A.

[0050] Figure 5(c) shows table T2c, which lists the top 10 words in order of frequency of occurrence as shown by submatrix U2c for each of topics 1 to 9 common to both companies, and Figure 5(b) shows table T2d, which lists the top 10 words in order of frequency of occurrence as shown by submatrix U2d for topics 10 to 18 specific to company B. By referring to these tables, users can understand the keywords that indicate the trends in company B's technology, separating them into topics common to company A and topics specific to company B.

[0051] Figure 6 shows graphs G1 and G2 output to the display after executing the document dataset analysis method S1.

[0052] Figure 6(a) shows a stacked line graph illustrating the time evolution of the weights represented by matrix W1 for topics 1-9 common to both companies and topics 10-18 specific to Company A. By referring to this graph, users can gain a detailed understanding of how the importance of each topic changes over time in Company A.

[0053] Figure 6(b) shows a stacked line graph illustrating the time evolution of the weights represented by matrix W2 for topics 1-9 common to both companies and topics 10-18 specific to Company B. By referring to this graph, users can gain a detailed understanding of how the importance of each topic changes over time in Company B.

[0054] (Additional notes) The present invention is not limited to the embodiments described above, and various modifications are possible within the scope of the claims. Embodiments obtained by appropriately combining the technical means disclosed in the embodiments described above are also included in the technical scope of the present invention. [Explanation of Symbols]

[0055] 1. Document Dataset Analysis Device 11 processors 12 Primary Memory 13 Secondary memory 14 Input / Output Interfaces 15 bus S1 Document Dataset Analysis Method S11 Generation Process S12 Disassembly Process S13 First Output Processing S14 Second output processing

Claims

1. Equipped with at least one processor, The aforementioned processor, A generation process that generates a three-dimensional non-negative tensor X1 having the dimensions of document data, words, and time, from a document dataset D1 consisting of document data containing time information, and a three-dimensional non-negative tensor X2 having the dimensions of document data, words, and time, with the frequency of occurrence of each word in each document data corresponding to each time, from a document dataset D2 consisting of document data containing time information, The following decomposition process is performed to factorize a three-dimensional non-negative tensor X1 into a two-dimensional matrix U1 with the dimensions of topic and word, where the elements are the frequency of occurrence of each word in each topic; a two-dimensional matrix V1 with the dimensions of document data and topic, where the elements are the weights of each topic in each document data; and a two-dimensional matrix W1 with the dimensions of time and topic, where the elements are the weights of each topic at each time; and to factorize a three-dimensional non-negative tensor X2 into a two-dimensional matrix U2 with the dimensions of topic and word, where the elements are the frequency of occurrence of each word in each topic; a two-dimensional matrix V2 with the dimensions of document data and topic, where the elements are the weights of each topic in each document data; and a two-dimensional matrix W2 with the dimensions of time and topic, where the elements are the weights of each topic at each time. The two-dimensional matrix U1 is composed of a submatrix U1c corresponding to topics common to document datasets D1 and D2, and a submatrix U1d corresponding to topics specific to document dataset D1. The two-dimensional matrix U2 is composed of a submatrix U2c corresponding to topics common to both document datasets D1 and D2, and a submatrix U2d corresponding to topics specific to document dataset D2. A document dataset analysis device characterized by the following features.

2. In the decomposition process, the processor determines the values ​​of the elements of the two-dimensional matrices U1, U2, V1, V2, W1, W2 in such a way that it minimizes a loss function in which the value increases as the submatrix U1c and submatrix U2c become more similar, and the value decreases as the submatrix U1d and submatrix U2d become more similar. The document dataset analysis device according to claim 1.

3. The processor further performs output processing to output a table T1c in which words are arranged in the order of frequency of occurrence shown by submatrix U1c for each topic common to document datasets D1 and D2, a table T1d in which words are arranged in the order of frequency of occurrence shown by submatrix U1d for each topic unique to document dataset D1, a table T2c in which words are arranged in the order of application frequency shown by submatrix U2c for each topic common to document datasets D1 and D2, and a table T2d in which words are arranged in the order of frequency of occurrence shown by submatrix U2d for each topic unique to document dataset D2. The document dataset analysis device according to claim 1.

4. The processor further performs output processing to output a graph G1 representing the time change of the weights shown by matrix W1 for each topic common to document datasets D1 and D2, and each topic specific to document dataset D1, and a graph G2 representing the time change of the weights shown by matrix W2 for each topic common to document datasets D1 and D2, and each topic specific to document dataset D2. The document dataset analysis device according to claim 1.

5. At least one processor generates a three-dimensional non-negative tensor X1 from a document dataset D1 consisting of document data containing time information, having dimensions of document data, words, and time, with elements being the frequency of occurrence of each word in each document data corresponding to each time, and generates a three-dimensional non-negative tensor X2 from a document dataset D2 consisting of document data containing time information, having dimensions of document data, words, and time, with elements being the frequency of occurrence of each word in each document data corresponding to each time, The processor includes a decomposition process that factors a three-dimensional non-negative tensor X1 into a two-dimensional matrix U1 having the dimensions of topic and word, with elements being the frequency of occurrence of each word in each topic; a two-dimensional matrix V1 having the dimensions of document data and topic, with elements being the weight of each topic in each document data; and a two-dimensional matrix W1 having the dimensions of time and topic, with elements being the weight of each topic at each time; and also factors a three-dimensional non-negative tensor X2 into a two-dimensional matrix U2 having the dimensions of topic and word, with elements being the frequency of occurrence of each word in each topic; a two-dimensional matrix V2 having the dimensions of document data and topic, with elements being the weight of each topic in each document data; and a two-dimensional matrix W2 having the dimensions of time and topic, with elements being the weight of each topic at each time. The two-dimensional matrix U1 is composed of a submatrix U1c corresponding to topics common to document datasets D1 and D2, and a submatrix U1d corresponding to topics specific to document dataset D1. The two-dimensional matrix U2 is composed of a submatrix U2c corresponding to topics common to both document datasets D1 and D2, and a submatrix U2d corresponding to topics specific to document dataset D2. A method for analyzing document datasets, characterized by the following features.

6. A program for operating a computer as a document dataset analysis device according to claim 1 or 2, wherein the computer is made to perform the generation process and the decomposition process. A document dataset analysis program characterized by the following features.

Images