Information processing method, program, and information processing apparatus

Abstract

This document provides an information processing method that can calculate explainable document features. [Solution] The information processing method involves acquiring multiple documents, extracting the words that make up each document, converting each extracted word into an embedding vector, classifying each word into multiple clusters based on the embedding vector, and then having a computer perform a process to calculate the document features of each document based on the occurrence frequency, which is the number of times each word classified into each cluster appears in a document, and the inverse document frequency, which is the reciprocal of the document frequency, which is the number of documents in which each word classified into each cluster appears, counted on a cluster-by-cluster basis.

Description

[[Technical Field]]

[0001] The present invention relates to an information processing method, a program, and an information processing apparatus. [[Background Art]]

[0002] In analyzing documents, it is common practice to calculate document feature quantities representing the content of documents (see, for example, Patent Documents 1 and 2). [[Prior Art Documents]] [[Patent Documents]]

[0003] [[Patent Document 1]] Japanese Unexamined Patent Application Publication No. 2019-159920 [[Patent Document 2]] Japanese Patent No. 7324577 [[Summary of the Invention]] [[Problem to be Solved by the Invention]]

[0004] In one aspect, an object of the present invention is to provide an information processing method or the like that can calculate an explainable document feature quantity. [[Means for Solving the Problem]]

[0005] The information processing method is executed by a computer that performs processing of: acquiring a plurality of documents; extracting words configuring each document; converting each extracted word into an embedding vector; classifying each word into a plurality of clusters based on the embedding vector; and calculating a document feature quantity for each document based on an occurrence frequency obtained by counting, for each cluster unit, the number of times words classified into each cluster appear in the document, and an inverse document frequency representing the reciprocal of document frequency obtained by counting, for each cluster unit, the number of documents in which words classified into each cluster appear. [[Effect of the Invention]]

[0006] In one aspect, an explainable document feature quantity can be calculated. [[Brief Description of the Drawings]]

[0007] [Figure 1] This is a block diagram showing an example of the configuration of an analysis device. [Figure 2] This figure shows an example of document analysis results. [Figure 3] This figure shows another example of document analysis results. [Figure 4] This flowchart shows an example of the processing procedure performed by the analysis device. [Modes for carrying out the invention]

[0008] The present invention will be described in detail below with reference to the drawings illustrating its embodiments. (Embodiment) Figure 1 is a block diagram showing an example configuration of the analysis device 1. In this embodiment, we will describe an analysis device (information processing device) 1 that calculates document features that balance discriminability (the ability to distinguish between different documents) and explainability by applying clustering and handcrafted feature extraction methods to an embedding model.

[0009] The analysis device 1 is an information processing device capable of various information processing functions, such as a personal computer or a server computer. In this embodiment, the analysis device 1 extracts words from a group of documents to be analyzed, converts each word into an embedding vector, and calculates document features for each document from the embedding vectors. Specifically, as will be described later, the analysis device 1 clusters the embedding vectors of each word and calculates document features for each document based on the frequency of occurrence of words classified into each cluster in the document and the inverse document frequency, which represents the reciprocal of the document frequency in which words classified into each cluster appear.

[0010] The analysis device 1 comprises a control unit 11, a main memory unit 12, a communication unit 13, a display unit 14, an input unit 15, and an auxiliary memory unit 16.

[0011] The control unit 11 is a processor such as one or more CPUs (Central Processing Units), MPUs (Micro-Processing Units), or GPUs (Graphics Processing Units), and performs various information processing by reading and executing programs P stored in the auxiliary storage unit 16. The main memory unit 12 is a temporary storage area such as SRAM (Static Random Access Memory) or DRAM (Dynamic Random Access Memory), and temporarily stores data necessary for the control unit 11 to perform calculations. The communication unit 13 is a communication module for performing communication-related processing, and transmits and receives information with the outside. The display unit 14 is a display screen such as an organic EL (Electro-Luminescence) display, and displays images. The input unit 15 is an operation interface such as a keyboard or mouse, and accepts operation input. The auxiliary storage unit 16 is a non-volatile storage area such as a hard disk, and stores programs P (program products) and other data necessary for the control unit 11 to perform processing.

[0012] The analysis device 1 may also be equipped with a reading unit that reads portable storage media 1a such as CD (Compact Disk)-ROM or DVD (Digital Versatile Disc)-ROM, and may read and execute the program P from the portable storage media 1a.

[0013] Figure 2 shows an example of the document analysis results. The outline of this embodiment will be described below.

[0014] Document features are calculated to enable document retrieval, classification, clustering, and visualization. i (i=1,2,…,N d ;N d Vector d represents the content of the document (number of documents). i It is defined as follows.

[0015] For example, in recent years, methods based on neural networks called embedding models have been used in various scenarios due to their high discriminative performance. However, there is a problem that document feature values obtained using only an embedding model do not have any meaning per se and are not explainable. In the present specification, a document feature value d i is defined as explainable when the words corresponding to each dimension thereof can be specified. Explainability of document feature values is particularly required when results of document search, classification, clustering, visualization, and the like are used for human decision-making.

[0016] In the present embodiment, newly introducing clustering and a handcrafted feature extraction method to an embedding model makes it possible to calculate explainable document feature values while maintaining high discriminative performance.

[0017] First, an analysis apparatus 1 performs morphological analysis on each document d to be analyzed i to obtain the document d i that includes constituent words t i j (j=1,2,…,T i ;T i is the number of words constituting d i and extracts the words. Next, the analysis apparatus 1 applies an embedding model to each word t i j to convert the word t i j into an embedding vector t i j Generally, t i j is learned such that the distance between words having a high co-occurrence frequency of surrounding words becomes small. Here, any embedding model such as BERT (Bidirectional Encoder Representations from Transformer) or Word2Vec can be employed as the embedding model. Note that in known methods, a document feature value d i is calculated, for example, by the following mathematical formula (1), which has the problem of not being explainable.

[0018]

number

[0019] In this embodiment, by introducing clustering and handcrafted feature extraction methods, explainable document features d i The following is calculated. First, the analysis device 1 calculates the set of embedding vectors T={t} transformed above. i j |i=1,2,…,N d ,j=1,2,…,T i Cluster the words} as shown in the following equation (2), and each word t i j Classify them into multiple clusters.

[0020]

number

[0021] Here, D d is the total number of clusters. The clustering method is not particularly limited, but for example, analysis device 1 performs clustering using the k-means method.

[0022] Furthermore, the analysis device 1 may exclude clusters from the calculation of document features described below, depending on the number of words included in the cluster and the embedding vectors of each word included in the cluster. Specifically, the analysis device 1 may exclude clusters that meet criteria such as having fewer than or equal to T1 words, or having a sum of squared residuals within the cluster of the embedding vectors of each word greater than or equal to T2, from subsequent processing. This makes it possible to select clusters that contribute to improved discriminability.

[0023] Analysis device 1 introduces a handcrafted feature extraction method, such as TF-IDF (Term Frequency-Inverse Document Frequency), to analyze each document d i Document features d representing the content i To calculate this, we define the set of all documents as D = {d1,d2,…,d NdIf expressed as}, then the document feature d i The kth element d i (k) is defined by the following equation (3):

[0024]

number

[0025] Here, tf(i,k) is document d i Among the words that make up cluster l k The frequency of occurrence of words belonging to (i.e., l k =f(t i j )(j=1,2,…,T i ) satisfying t i j The number of (i,k) is such that tf(i,k) is cluster l k Words classified as documents d i This is the frequency of occurrence, counted not at the word level, but at the cluster level. tf(i,k) corresponds to the TF value (Term Frequency) in TF-IDF, but differs from the TF value in that it is counted at the cluster level, not at the word level. For example, word t i 1 and the word t i 2 If and are classified into cluster l1, then in TF-IDF, word t i 1 t i 2 Document d for each i The number of times a word appears is calculated as the TF value, but in this embodiment, the two words t i 1 , t i 2 document d i The total number of occurrences is calculated as tf(i,k).

[0026] df(k) is cluster l k This is the total number of documents in which words belonging to cluster l appear. In other words, df(k) is the total number of documents in cluster l kThe document frequency is calculated by counting the number of documents in which a word classified as such appears, on a cluster basis. log(|D| / df(k)) is the inverse document frequency, which is the reciprocal of that document frequency. log(|D| / df(k)) is equivalent to the IDF value (Inverse Document Frequency) in TF-IDF, but differs from the IDF value in that it is counted on a cluster basis, not on a word basis.

[0027] The document feature d is defined by the above formula (3). i This is designed so that the distance between document features of documents that have similar topics is small, and it has high discriminative ability. In addition, the document features d obtained by formula (3) i Each dimension (the kth element) is the kth cluster l k It supports cluster l k It is clear that this is a set of words classified by formula (2). Therefore, the document feature d i Unlike document features obtained using only embedding models, this is explainable because the words corresponding to each dimension are clear.

[0028] The obtained document features d i This is a multidimensional vector and can be applied to various tasks. In this embodiment, as an example, we will describe its application to a visualization task that visualizes document features and a clustering task that clusters documents. In this embodiment, we performed analysis on labor accident reports related to the power industry.

[0029] First, let's explain the visualization task. Analysis device 1 performs principal component analysis as shown in equation (4) below, and extracts document features d i This is projected onto a feature space of a predetermined number of dimensions (2 dimensions in this embodiment).

[0030]

number

[0031] U is a matrix of eigenvectors obtained by the eigenvalue decomposition shown in the following equation (5).

[0032]

number

[0033] Here, D is the document feature d i This is a matrix centered by arranging the elements. Based on the size of the elements of U, the clusters contributing to the principal component axes are l k And the words to which it belongs can be identified. Also, the (k, l)th element of U (l=1,2) is cluster l k This represents the contribution of words belonging to this group to the first principal component axis.

[0034] The analysis device 1 generates a map by projecting each document onto the feature space as described above, and displays it on the display unit 14. Figure 2 shows an example of the map display. Each axis on the map represents a principal component, and each dot on the map represents a document.

[0035] The analysis device 1 displays the word groups belonging to the clusters that contributed to each principal component, corresponding to each principal component axis. For example, as shown in Figure 2, the analysis device 1 displays the word groups in a word cloud format. The analysis device 1 identifies the clusters that contributed to each principal component based on the eigenvectors calculated during the principal component analysis, and displays the word groups belonging to the identified clusters, corresponding to each principal component axis of the map.

[0036] Furthermore, the analysis device 1 changes the display method of words belonging to each cluster according to the contribution of each cluster to each principal component. Specifically, the analysis device 1 increases the font size of words belonging to clusters with high contributions. The analysis device 1 identifies the contribution of each cluster to each principal component based on the eigenvectors calculated during principal component analysis, and displays a word cloud with the size of words belonging to each cluster changed according to the identified contribution.

[0037] Thus, the analysis device 1 displays a map that projects each document onto the feature space, showing the group of words classified into clusters that contributed to each principal component, displayed in a manner corresponding to the degree of contribution of each cluster. This not only allows for the visualization of the document features of each document, but also allows for the presentation (explanation) of the group of words that contributed to each principal component axis. In contrast, when principal component analysis is performed on document features obtained using known methods such as Doc2Vec, it is possible to visualize the document features of each document, but it is not possible to reveal the group of words that contributed to each principal component axis.

[0038] Next, we will explain the clustering task. The analysis device 1 analyzes the document features d of each document. i Clustering is performed on the documents, and the attributes of each document are estimated by classifying them into multiple groups. The clustering method is not particularly limited, but for example, analysis device 1 performs clustering using the k-means method.

[0039] Figure 3 shows another example of document analysis results. Figure 3 illustrates a feature map showing the clustering results. The dotted lines in Figure 3 indicate groups of classified documents. For example, analysis device 1 displays a map that clearly identifies documents belonging to each group by changing the display color of the dots representing each document for each group. "Bee sting," "electric shock," and "cut" shown in Figure 3 are accident events reported in documents (work accident reports) belonging to each group, and represent the group name (attribute) of each group.

[0040] Group names may be assigned manually, but they can also be assigned automatically by analyzing the documents belonging to each group. For example, the analysis device 1 assigns the most frequently occurring words in the document group to each group as the group name. Alternatively, the analysis device 1 may input the document group belonging to each group into a large-scale language model to generate a summary, and then assign the words in the generated summary as the group name. The analysis device 1 displays the group names assigned to each group, associating them with the respective groups.

[0041] Furthermore, Table 1 below shows the quantitative evaluation metrics for the document clustering results. In this embodiment, quantitative evaluation was performed using the BBC News Archive, a publicly available dataset containing 2,225 documents and defined ground truth clusters for each document. Table 1 shows the NMI (Normalized Mutual Information) and explainability metrics for clustering for the well-known Doc2Vec, TF-IDF, and this method. From Table 1, it was confirmed that both discriminability and explainability were achieved.

[0042] [Table 1]

[0043] As shown in Table 1, this method has a lower NMI than the neural network-based method (Doc2Vec), but the neural network-based method sacrifices explainability. Furthermore, this method has a higher NMI than the frequency analysis-based method (TF-IDF) and exhibits higher discriminative ability than the frequency analysis-based method.

[0044] Figure 4 is a flowchart showing an example of the processing procedure performed by the analysis device 1. Based on Figure 4, the processing content performed by the analysis device 1 will be explained. The control unit 11 of the analysis device 1 acquires multiple documents (step S11). The control unit 11 extracts the words that make up each document (step S12). The control unit 11 converts each extracted word into an embedding vector (step S13).

[0045] The control unit 11 classifies each word into multiple clusters based on the transformed embedding vector (step S14). The control unit 11 calculates document features for each document based on the occurrence frequency, which is the number of times each word classified into each cluster appears in a document, counted on a cluster-by-cluster basis, and the inverse document frequency, which is the reciprocal of the document frequency, which is the number of documents in which each word classified into each cluster appears, counted on a cluster-by-cluster basis (step S15).

[0046] The control unit 11 projects each document onto a feature space with a predetermined number (e.g., 2) principal components by performing principal component analysis on the document features of each document (step S16). Based on the eigenvectors calculated during the principal component analysis, the control unit 11 identifies the clusters that contributed to each principal component (step S17). Specifically, the control unit 11 identifies the clusters that contributed to each principal component and the degree of contribution of each cluster to each principal component.

[0047] The control unit 11 classifies each document into multiple groups according to its document features (step S18). Based on the documents belonging to each group, the control unit 11 assigns a group name to each group (step S19). For example, the control unit 11 may assign the most frequent word in the documents belonging to each group as the group name, or it may generate a summary of the documents belonging to each group and assign a word in the generated summary as the group name.

[0048] The control unit 11 displays a map projecting each document onto the feature space (step S20). Specifically, based on the identification results in step S17, the control unit 11 displays a map showing the groups of words classified into clusters contributing to each principal component. In this case, the control unit 11 changes the display manner (e.g., size) of each word according to the contribution of each cluster to the principal component. Also in step S20, based on the classification results in step S18, the control unit 11 displays a map that makes it possible to identify the documents belonging to each group. In this case, the control unit 11 displays the group name assigned to each group in association with each group. The control unit 11 then completes the series of processes.

[0049] Based on the above, according to this embodiment, explainable document features can be calculated.

[0050] The embodiments disclosed herein should be considered in all respects to be illustrative and not restrictive. The scope of the present invention is indicated by the claims, not in the sense described above, and all modifications within the sense and scope equivalent to the claims are intended.

[0051] The matters described in each embodiment can be combined with each other. Furthermore, the independent and dependent claims described in the claims can be combined with each other in any combination, regardless of the form of reference. In addition, the claims use a form in which claims referencing two or more other claims (multi-claim form), but are not limited to this. A form in which multi-claims referencing at least one multi-claim (multi-multi-claim) may also be used. [Explanation of Symbols]

[0052] 1. Analysis device (information processing device) 11 Control Unit 12 Main memory 13 Communications Department 14 Display section 15 Input section 16 Auxiliary storage P Program

Claims

1. Retrieve multiple documents, Extract the words that make up each document, Convert each extracted word into an embedding vector, Based on the aforementioned embedding vector, each word is classified into multiple clusters, The document features of each document are calculated based on the frequency of occurrence (counting the number of times each word classified into each cluster appears in a document, on a cluster-by-cluster basis) and the inverse document frequency (representing the reciprocal of the document frequency, which is the number of documents in which each word classified into each cluster appears, on a cluster-by-cluster basis). An information processing method in which a computer performs the processing.

2. The total number of clusters is D ｄ If D is the set of all documents, tf(i,k) is the frequency of occurrence of a word classified into cluster k in the i-th document, and df(k) is the total number of documents in which a word classified into cluster k appears, then the k-th element d of the document features of the i-th document is defined as follows: ｉ (k) is calculated according to the following formula (3). The information processing method according to claim 1. [Math 1]

3. Depending on the number of words in a cluster, or the embedding vectors of each word in the cluster, the cluster is excluded from the calculation of document features. The information processing method according to claim 1.

4. By performing principal component analysis on the document features of each document, each document is projected onto a feature space with a predetermined number of principal components as axes. Based on the eigenvectors calculated during principal component analysis, the clusters contributing to each principal component are identified. The output map projects each document onto the feature space and shows the group of words classified into the clusters that contributed to each principal component. The information processing method according to claim 1.

5. Based on the aforementioned eigenvectors, the contribution of each cluster to each principal component is identified. The map is output, with the display of each word changed according to the aforementioned contribution. The information processing method according to claim 4.

6. Each document is classified into multiple groups according to its document features. Output the map that displays the documents belonging to each group in an identifiable manner. The information processing method according to claim 4.

7. A group name is assigned to each group based on the documents belonging to that group. The information processing method according to claim 6.

8. Obtain an embedding vector by transforming each word that makes up each of the multiple documents. Based on the aforementioned embedding vector, each word is classified into multiple clusters, The document features of each document are calculated based on the frequency of occurrence (counting the number of times each word classified into each cluster appears in a document, on a cluster-by-cluster basis) and the inverse document frequency (representing the reciprocal of the document frequency, which is the number of documents in which each word classified into each cluster appears, on a cluster-by-cluster basis). A program that instructs a computer to perform a process.

9. An information processing device comprising a control unit, The control unit, Retrieve multiple documents, Extract the words that make up each document, Convert each extracted word into an embedding vector, Based on the aforementioned embedding vector, each word is classified into multiple clusters, The document features of each document are calculated based on the frequency of occurrence (counting the number of times each word classified into each cluster appears in a document, on a cluster-by-cluster basis) and the inverse document frequency (representing the reciprocal of the document frequency, which is the number of documents in which each word classified into each cluster appears, on a cluster-by-cluster basis). Information processing device.

Images