JP2024165421A - Information processing device, information processing method, and information processing program - Google Patents

Abstract

[Problem] The importance of each sentence included in a document is obtained while taking into consideration the logical structure of the document. [Solution] The present invention relates to an information processing device. The information processing device of the present invention is characterized by having a graph document storage means for storing a graph document in which an input document including a plurality of sentences is converted into a graph format, an important sentence extraction means for extracting one or a plurality of sentences as important sentences from the graph document, and a summary generation means for generating a summary that summarizes the input document based on the sentences extracted as important sentences by the important sentence extraction means. [Selected Figure] Figure 1

Description

This invention relates to an information processing device, an information processing method, and an information processing program, and can be applied, for example, to the process of summarizing a document.

Conventionally, there are two main types of methods for automatically summarizing documents (data containing sentences): "extraction-type summarization" and "abstraction-type summarization." Extraction-type summarization is a method in which sentences of high importance are extracted from the sentences in a document and then combined to create a summary. Abstraction-type summarization is a method in which a summary is generated by inputting the document to be summarized into a generative model such as a neural network. While abstraction-type summarization makes it easy to control the length of the summary, it can sometimes generate a summary that differs from the content of the input document. In other words, extraction-type summarization is preferable in situations where an accurate summary with no errors in the information is required.

Incidentally, conventional extraction-based summarization, as disclosed in Patent Document 1 and Non-Patent Document 1, has used a method of creating a graph based on the similarity between sentences contained in a document and calculating the importance of each sentence from the graph structure. For example, the automatic summarization method described in Patent Document 1 calculates the importance of each sentence using not only the similarity between sentences contained in a document, but also the relevance between each sentence in the document and keywords that reflect the user's preferences.

JP 2020-57438 A

Erkan, Guenes, and Dragomir R. Radev. "Lexrank: Graph-based lexical centrality as salience in text summarization." Journal of artificial intelligence research 22 (2004): 457-479.

However, the automatic summarization method described in Patent Document 1 has the problem that it ignores the order in which each sentence appears and the relationships between sentences when calculating the importance of each sentence in a document, and does not take into account the logical structure of the document.

In view of the above problems, there is a need for an information processing device, an information processing method, and an information processing program that obtain the importance of each sentence contained in a document while taking into account the logical structure of the document.

The information processing device of the first invention is characterized by having a graph document storage means for storing a graph document in which an input document including a plurality of sentences is converted into a graph format, an important sentence extraction means for extracting one or more sentences as important sentences from the graph document, and a summary generation means for generating a summary that summarizes the input document based on the sentences extracted as important sentences by the important sentence extraction means.

The second information processing program of the present invention is characterized in that it causes a computer to function as a graph document storage means for storing a graph document in which an input document containing a plurality of sentences is converted into a graph format, an important sentence extraction means for extracting one or more sentences as important sentences from the graph document, and a summary generation means for generating a summary that summarizes the input document based on the sentences extracted as important sentences by the important sentence extraction means.

The third invention is an information processing method performed by an information processing device, characterized in that the information processing device has a graph document storage means, an important sentence extraction means, and a summary generation means, the graph document storage means stores a graph document in which an input document including a plurality of sentences is converted into a graph format, the important sentence extraction means extracts one or a plurality of sentences as important sentences from the graph document, and the summary generation means generates a summary that summarizes the input document based on the sentences extracted as important sentences by the important sentence extraction means.

In the first aspect of the present invention, the system may further include a graph generating means for generating the graph document based on the input document when the input document is input, and the graph document storage means may store the graph document generated by the graph generating means.

Furthermore, in the first aspect of the present invention, the graph generating means may be characterized in that it divides the input document into sentences of a predetermined unit, obtains a collection of the divided sentences, and generates, as the graph document, data having a graph structure in which each sentence constituting the obtained collection of sentences is a node and the logical relationships between each node are edges.

Furthermore, in the first aspect of the present invention, the important sentence extraction means may generate a logical structure matrix in which the logical structure expressed in the graph format in the graph document is expressed in the form of a matrix, obtain the importance of sentences corresponding to each node constituting the graph document from eigenvectors based on the logical structure matrix, and extract one or more sentences as important sentences from the collection of sentences based on the obtained importance.

In addition, in the first aspect of the present invention, the key sentence extraction means may perform a numerical conversion process to convert the logical relationships of the edges constituting the graph document into a numerical expression to generate a converted graph document, and generate the logical structure matrix based on the converted graph document.

Furthermore, in the first aspect of the present invention, the key sentence extraction means may perform the numerical conversion process based on a machine-learned learning model.

According to the present invention, it is possible to obtain the importance of each sentence contained in a document while taking into account the logical structure of the document.

1 is a block diagram showing a functional configuration of a document summarization device according to an embodiment. FIG. 2 is a diagram showing a structure of a graph document according to the embodiment. 4 is a flowchart showing the operation of the document summarizing device according to the embodiment. 11 is a flowchart showing a specific example of an important sentence extraction process performed by an important sentence extraction unit according to the embodiment. 4 is a diagram showing an example of the configuration of a numerical value conversion table used in the document summarization device according to the embodiment; FIG. 11 is a diagram showing an example of a state in which a numeric conversion process is performed on the labels (logical relationships) of the edges of a graph document used in the document summarization device according to the embodiment. FIG.

An embodiment of an information processing device, an information processing method, and an information processing program according to the present invention will be described in detail below with reference to the drawings. In this embodiment, an example in which the information processing device, the information processing method, and the information processing program according to the present invention are applied to a document summarization device will be described.

(A-1) Configuration of the Embodiment FIG. 1 is a block diagram showing the functional configuration of a document summarization device 10 according to this embodiment.

The document summarization device 10 is a device that performs document summarization processing to summarize input data including a document composed of multiple sentences, and outputs output data including the summary. Hereinafter, the document included in the input data will be referred to as the "input document." Note that the summary includes one or more sentences.

The document summarization device 10 has a document input unit 11, a preprocessing unit 12, a key sentence extraction unit 13, a summary generation unit 14, and a summary output unit 15.

The document summarization device 10 may be configured entirely from hardware (e.g., a dedicated semiconductor chip, etc.), or may be configured partially or entirely from a computer program (software). For example, the document summarization device 10 may be realized by installing a program (including the "information processing program" according to the embodiment) in a computer (not shown) equipped with a memory, a processor, etc.

The document input unit 11 performs a process of accepting input data (input document). The means by which the document input unit 11 receives input data is not limited, and various means can be applied. For example, the document input unit 11 may receive input data from another device via communication, may receive data offline from a data recording medium (e.g., a data recording medium such as a USB memory or a hard disk), or may receive manual input from a user (e.g., input using an input device such as a keyboard not shown).

The preprocessing unit 12 is responsible for preprocessing (details will be described later) the documents contained in the input data. The preprocessing unit 12 has a document division unit 121 and a graph generation unit 122.

The document division unit 121 performs a process of dividing an input document into sentences. Hereinafter, each sentence divided from an input document is referred to as a "divided sentence". The document division unit 121 acquires a plurality of divided sentences from the input document and outputs a collection of the acquired divided sentences. The method by which the document division unit 121 divides an input document is not limited, and various methods can be used. The document division unit 121 may extract divided sentences according to a predetermined algorithm (for example, extracting divided sentences based on symbols such as punctuation marks and line break codes contained in the input document), or may extract divided sentences using a machine learning model that has previously learned pairs of "plain text" and "collections of divided sentences" (collections of divided sentences that are the correct answer) as training data.

The graph generation unit 122 uses the collection of divided sentences acquired by the document division unit 121 to perform processing to generate a document in graph format (graph structure) (hereinafter referred to as a "graph document").

Figure 2 shows the structure of a graph document in this embodiment.

Figure 2 shows an example of a graph document structure based on the plain text input document "It will rain tomorrow. The field trip will be canceled. The probability of precipitation is 90%."

As shown in FIG. 2, in the graph document of this embodiment, each divided sentence is assigned to one node, and each edge indicates the logical relationship between the nodes at both ends (between divided sentences). In FIG. 2, each node is shown with a rectangular symbol, and the edges between the nodes are illustrated with arrows. In the graph document of this embodiment, each edge is assigned a directionality according to the direction of the arrow. In FIG. 2, a divided sentence is attached to the symbol of each node, and a label indicating the logical relationship is attached to the arrow of each edge.

Examples of logical relationships corresponding to edges (logical relationships between two divided sentences) include patterns such as "reason," "addendum," and "example." In this embodiment, the specific data format of the graph document is not limited, and data structures such as databases (middleware that constitutes a database) in various graph formats can be applied.

In the example of the graph document structure shown in FIG. 2, three divided sentences, "It will rain tomorrow," "The field trip will be canceled," and "The probability of precipitation is 90%," are extracted from the above input document (plain text), and nodes X1, X2, and X3 are assigned to each divided sentence. In addition, in the example of the graph document structure shown in FIG. 2, an edge E1 labeled "Result" (logical relationship) is set from node X1 (It will rain tomorrow) to node X2 (The field trip will be canceled). Furthermore, in the example of the graph document structure shown in FIG. 2, an edge E2 labeled "Supplement" (logical relationship) is set from node X1 (It will rain tomorrow) to node X3 (The probability of precipitation is 90%).

As described above, in this embodiment, the graph document is assumed to have a structure as shown in the example of Figure 2.

The specific processing method of the graph generation unit 122 for generating a graph document is not limited. For example, the graph generation unit 122 may be configured to use a machine learning model that has been trained using pairs of a "collection of divided sentences" and a corresponding "graph document" (a graph document that is the correct answer) as training data, but is not limited to this. Note that in this embodiment, the graph generation unit 122 generates a graph document from a collection of divided sentences, but the graph document may also be generated directly from plain text. For example, if the graph generation unit 122 is provided with a machine learning model that has been trained using pairs of a "plain text" and a corresponding "graph document" (a graph document that is the correct answer) as training data, it becomes possible to generate a graph document directly from plain text.

The important sentence extraction unit 13 performs a process of extracting important sentences from each node (split sentence) constituting the graph document supplied from the preprocessing unit 12. The important sentence extraction unit 13 calculates a numerical value (hereinafter referred to as an "importance value") indicating the degree of importance for each node (split sentence) of the graph document, and extracts one or more split sentences as important sentences based on this importance value. For example, the important sentence extraction unit 13 may extract split sentences with high importance (for example, a predetermined number of split sentences from the top) as important sentences from the graph document.

The summary output unit 15 performs a process of generating a summary based on the important sentences (one or more divided sentences) extracted by the important sentence extraction unit 13. The summary output unit 15 also outputs data including at least the summary as output data. The summary output unit 15 may treat the important sentences as they are as summaries, or may treat the important sentences further processed (e.g., summarization processing using another document summarization algorithm) as summaries. The method by which the summary output unit 15 outputs output data is not limited. For example, the summary output unit 15 may transmit the output data to another device via communication, may write the output data offline to a data recording medium (e.g., a data recording medium such as a USB memory or a hard disk), may display the output data on a display device such as a display, or may print the output data from a printing device such as a printer.

(A-2) Operation of the Embodiment Next, the operation of the document summarization device 10 of this embodiment (information processing method according to the embodiment) will be described.

Figure 3 is a flowchart showing the operation of the document summarization device 10.

First, it is assumed here that input data including an input document is input to the document summarization device 10 (S101). In the document summarization device 10, the input data (input document) is acquired by the document input unit 11. The document input unit 11 supplies the input document included in the input data to the preprocessing unit 12.

When the preprocessing unit 12 receives an input document, it first analyzes the data structure of the input document (S102). Here, the input document is assumed to be either a graph document (a document having structural information in a graph format/graph structure) or plain text without any particular structural information. In other words, the preprocessing unit 12 in this embodiment analyzes the structure of the input document to determine whether it is a graph document having structural information in a graph format or plain text without any particular structural information. If the input document is a graph document, the preprocessing unit 12 supplies the input document as is to the key sentence extraction unit 13 and proceeds to step S105, which will be described later, and if the input document is plain text, the preprocessing unit 12 proceeds to step S103, which will be described later, and performs processing to convert the input document into a graph document.

If the input document is plain text, the preprocessing unit 12 first supplies the input document (plain text) to the document division unit 121, which divides the input document into sentences (divided sentences) (S103).

Next, the preprocessing unit 12 supplies the set of divided sentences divided by the document division unit 121 to the graph generation unit 122 to generate a graph document (S104), and supplies the generated graph document to the key sentence extraction unit 13.

Through the above processing, the preprocessing unit 12 acquires the graph document and supplies it to the key sentence extraction unit 13.

Next, the important sentence extraction unit 13 calculates an importance value for each node (each divided sentence) of the supplied graph document, performs a process of extracting divided sentences (nodes) that are important sentences based on the calculation result (hereinafter referred to as "important sentence extraction process") (S105), and supplies the extracted important sentences (one or more divided sentences) to the summary generation unit 14. At this time, the number of divided sentences extracted as important sentences is not limited. For example, the important sentence extraction unit 13 may be able to arbitrarily set the number of divided sentences to be extracted as important sentences by a user operation. The format of the data supplied from the important sentence extraction unit 13 to the summary output unit 15 is not limited. For example, the important sentence extraction unit 13 may supply each divided sentence extracted as an important sentence to the summary generation unit 14 as plain text, or may supply a list of identifiers of the nodes (divided sentences) extracted as important sentences together with the graph document to the summary generation unit 14.

Next, the summary generation unit 14 creates a summary from the supplied important sentences (S106) and supplies it to the summary output unit 15. The format of the summary is not limited. For example, the summary may be in plain text format or in graph format. For example, the summary generation unit 14 may acquire plain text in which important sentences (plain text of divided sentences) are linked as a summary. In addition, for example, the summary generation unit 14 may acquire a summary in which nodes (divided sentences) corresponding to important sentences are highlighted (indicated as important sentences) in the graph document. Specifically, for example, the summary generation unit 14 may acquire a summary in which information (for example, information such as a label) indicating that the nodes (divided sentences) corresponding to important sentences are important sentences is added to the nodes (divided sentences) corresponding to important sentences in the graph document.

Finally, the summary output unit 15 performs processing to output the provided summary in a predetermined format (S107).

As described above, the document summarization device 10 performs the process of creating a summary corresponding to the input document.

Next, the details of the process of step S105 described above in the document summarization device 10 (the key sentence extraction process by the key sentence extraction unit 13) will be described.

Figure 4 is a flowchart showing a specific example of the key sentence extraction process performed by the key sentence extraction unit 13.

First, it is assumed that a graph document to be subjected to the key sentence extraction process is input to the key sentence extraction unit 13 (S201). Here, the following explanation will be given assuming that a graph document having the structure shown in FIG. 2 is input to the key sentence extraction unit 13.

The key sentence extraction unit 13 first performs a process of converting the values of each edge in the input graph document (logical relationships set in the labels) into a numerical expression as a scalar (hereinafter referred to as the "numerical conversion process") (S202).

The specific method of the numerical conversion process is not limited, and may be performed, for example, using a preset table (hereinafter referred to as a "numerical conversion table"), or a machine-learned learning model (e.g., a neural network). For example, the important sentence extraction unit 13 may obtain, for each logical relationship (name of logical relationship), a value based on a word vector (e.g., a word vector when the name of the logical relationship is a word) generated using a machine-learned learning model as a numerically expressed value. The important sentence extraction unit 13 in this embodiment is provided with a numerical conversion table in advance, and performs numerical conversion process on the label of each edge using this numerical conversion table.

Figure 5 shows an example of the configuration of a numeric conversion table.

In FIG. 5, a corresponding numerical value is registered for each logical relationship (the name of the logical relationship). For example, in FIG. 5, numerical values are set for each logical relationship of result, reason, supplement, example, etc. The number and combination of logical relationships set in the numerical conversion table are not limited. Also, in FIG. 5, numerical values of 0.8, 0.5, 0.2, and 0.2 are set for result, reason, supplement, and example, respectively, but the numerical values corresponding to each logical relationship are not limited and various values may be set. In the numerical conversion table, for example, the numerical value may be set to a larger value for a logical relationship that tends to be more important when summarizing a document. For example, in the example of FIG. 5, it is assumed that sentences connected by the logical relationships of "result" and "reason" tend to be more important when summarizing, but sentences connected by the logical relationships of "reason" and "supplement" tend to be less important when summarizing compared to the above two logical relationships ("result" and "reason"). Furthermore, the example in Figure 5 is based on the premise that, when summarizing a document, sentences connected by the logical relationship of "result" tend to be relatively more important than sentences connected by the logical relationship of "reason." Based on the above premise, in Figure 5, values are set for each logical structure so that the magnitude relationship is "result > reason > supplement = example."

Figure 6 shows the state of the labels (logical relationships) of each edge in the graph document shown in Figure 2 after numerical conversion processing based on the numerical conversion table shown in Figure 5.

As shown in Figure 6, based on the numerical conversion table shown in Figure 5, the key sentence extraction unit 13 sets the numerical value for edge E1 whose label (logical relationship) is "result" to 0.8, and sets the numerical value for edge E2 whose label (logical relationship) is "supplement" to 0.2.

As described above, the key sentence extraction unit 13 performs a numerical conversion process on the labels (logical relationships) of each edge.

Next, the key sentence extraction unit 13 generates a logical structure matrix L that represents the logical structure between each node of the graph document that has been subjected to the numerical conversion process (S203).

Here, the logical structure matrix L of a graph document is defined as the product of each component (component corresponding to each edge) of the adjacency matrix that represents the connection relationships between each node and the numerical representation of the corresponding edge (numerical value corresponding to the logical relationship). The logical structure matrix L can be said to represent, in matrix form, the logical structure that is represented in graph form in the original graph document.

The following formula (1) represents the adjacency matrix A based on the graph document shown in Figure 6 (after numerical conversion processing for the logical relationships of the edges). In adjacency matrix A, the first to third rows correspond to nodes X1 to X3, respectively, and the first to third columns correspond to X1 to X3, respectively. Then, in adjacency matrix A, edge E1 directed from node X1 to node X2 is the element in the second column of the first row, and edge E2 directed from node X1 to node X3 is the element in the third column of the first row. In adjacency matrix A shown in formula (1), the elements corresponding to edges E1 and E2 have a value of 1, and the other elements are 0.

The following formula (2) expresses the logical structure matrix L based on the graph document shown in Fig. 6 (where the logical relationships of the edges have been subjected to numerical conversion processing). The logical structure matrix L is the result of multiplying each element (element corresponding to each edge) of the adjacency matrix A by the numerical expression of the corresponding edge (numerical value corresponding to the logical relationship). Therefore, in the logical structure matrix L of the following formula (2), the element (first row, second column) corresponding to edge E1 is "1 x 0.8 = 0.8", and the element (first row, third column) corresponding to edge E2 is "1 x 0.2 = 0.2".

Next, the key sentence extraction unit 13 adds the logical structure matrix L created in step S203 and the random jump matrix R at a constant ratio d (S204).

Here, the random jump matrix R is a matrix that has the same number of rows and columns as the logical structure matrix L, and all elements (components) are 1/N. Here, N is the total number of nodes in the original graph document. Here, the number of nodes in the graph document shown in Figure 6 is 3, so N = 3. Therefore, here, the random jump matrix R has the number of rows and columns N = 3, and all elements are 1/N = 1/3, as shown in the following formula (3).

Here, the important sentence extraction unit 13 obtains a matrix M by adding a logical structure matrix L and a random jump matrix R in a ratio of d:(1-d) as shown in the following formula (4). In formula (4), d=0.85 is used, but the value of d is not limited and may be set to any value (for example, a suitable value obtained by an experiment, a simulation, or the like).

Next, the key sentence extraction unit 13 finds the eigenvector x of the matrix M created in step S204 (S205).

The procedure for the important sentence extraction unit 13 to obtain the eigenvector is not limited, but for example, the following procedure can be applied. The important sentence extraction unit 13 first obtains a matrix M' by dividing the components of each row in the matrix M by the sum of the rows. As a result, the sum of each row in the matrix M' becomes 1. When the matrix M' is obtained based on the matrix M shown in the formula (4), the result becomes the formula (5). Next, the important sentence extraction unit 13 obtains an eigenvector of M' with an eigenvalue of 1 that satisfies M'x = λx as x. When the eigenvector x is obtained based on the matrix M' shown in the formula (5), the result becomes the following formula (6). The eigenvector x is a vector represented by a matrix of 1 row and N columns (N is the total number of nodes in the original graph document). In the eigenvector x, the elements of each column indicate the importance of the node (the divided sentence corresponding to the node) in the corresponding column. Note that the nodes corresponding to each column in the eigenvector x match the columns of the original logical structure matrix L (adjacency matrix A). For example, the eigenvector x shown in the following formula (5) has one row and three columns, and the first to third rows correspond to nodes X1 to X3, respectively. In the following formula (5), the importance corresponding to node X1 (it will rain tomorrow) is 0.26, the importance corresponding to node X2 (the field trip will be canceled) is 0.44, and the importance corresponding to node X3 (the probability of precipitation is 90%) is 0.30.

Next, the important sentence extraction unit 13 extracts the divided sentences corresponding to the nodes with high importance determined in step S205 as important sentences (S206).

Here, the important sentence extraction unit 13 extracts the divided sentence corresponding to the node with the highest importance as an important sentence. Then, when an eigenvector like that of equation (5) is obtained, the important sentence extraction unit 13 extracts the divided sentence "The field trip is canceled" corresponding to node X2 as an important sentence.

(A-3) Advantages of the Embodiment The document summarization device 10 of this embodiment can provide the following advantages.

In the document summarization device 10 of this embodiment, the structure of a graph document based on an input document (plain text) is expressed by a logical structure matrix L as described above, and an eigenvector x based on this logical structure matrix L is obtained, thereby making it possible to perform a process of extracting important sentences while taking into consideration the order in which sentences in the input document appear and the relationships between the sentences. In particular, in the graph document used in the document summarization device 10 of this embodiment, the logical relationships between nodes (sentences) are expressed using directional edges, so that an eigenvector x (importance of each divided sentence) that reflects the order in which sentences in the input document appear and the relationships between the sentences can be obtained. Therefore, in the document summarization device 10 of this embodiment, it becomes possible to accurately include important sentences in the input document in the summary.

(B) Other Embodiments The present invention is not limited to the above-described embodiments, and modified embodiments such as those exemplified below can also be mentioned.

(B-1) In the above embodiment, the document divider 121 divides the input document so that each divided sentence contains only one sentence, but the divided sentence may contain multiple sentences (for example, a collection of multiple sentences such as a paragraph).

(B-2) In the above embodiment, the input document included in the input data input to the document summarization device 10 is either plain text or a graph document, but it may be assumed that only one of them is input. For example, if the input data contains only plain text, the process of step S102 (document format determination process) may be omitted in the flowchart of FIG. 3, and steps S103 and S104 (preprocessing for converting plain text into a graph document) may be performed after step S101. Also, for example, if the input data contains only graph documents, the preprocessing unit 12 may be omitted in the document summarization device 10, and the processes of steps S102 to S104 (document format determination process and preprocessing) may be omitted in the flowchart of FIG. 3.

10... document summarization device, 11... document input unit, 12... preprocessing unit, 13... important sentence extraction unit, 14... summary generation unit, 15... summary output unit, 121... document segmentation unit, 122... graph generation unit

Claims (8)

A graph document storage means for storing a graph document in which an input document including a plurality of sentences is converted into a graph form;
an important sentence extraction means for extracting one or more sentences as important sentences from the graph document;
and a summary generating means for generating a summary that summarizes the input document based on the sentences extracted as key sentences by the key sentence extracting means. a graph generating unit that generates the graph document based on the input document when the input document is input;
2. The information processing apparatus according to claim 1, wherein said graph document storage means stores said graph document generated by said graph generating means. The information processing device according to claim 2, characterized in that the graph generation means divides the input document into sentences of a predetermined unit, acquires a collection of the divided sentences, and generates, as the graph document, data having a graph structure in which each sentence constituting the acquired collection of sentences is a node and the logical relationships between each node are edges. The information processing device according to claim 3, characterized in that the important sentence extraction means generates a logical structure matrix in which the logical structure expressed in the graph format in the graph document is expressed in the form of a matrix, obtains the importance of sentences corresponding to each node constituting the graph document from eigenvectors based on the logical structure matrix, and extracts one or more sentences as important sentences from the collection of sentences based on the obtained importance. The information processing device according to claim 4, characterized in that the key sentence extraction means performs a numerical conversion process to convert the logical relationships of each edge constituting the graph document into a numerical expression to generate a converted graph document, and generates the logical structure matrix based on the converted graph document. The information processing device according to claim 5, characterized in that the key sentence extraction means performs the numerical conversion process based on a machine-learned learning model. Computer,
A graph document storage means for storing a graph document in which an input document including a plurality of sentences is converted into a graph form;
an important sentence extraction means for extracting one or more sentences as important sentences from the graph document;
2. An information processing program that causes the program to function as a summary generating means for generating a summary that summarizes the input document based on the sentences extracted as key sentences by the key sentence extracting means. An information processing method performed by an information processing device,
The information processing device includes a graph document storage means, an important sentence extraction means, and a summary generation means,
The graph document storage means stores a graph document in which an input document including a plurality of sentences is converted into a graph format;
The key sentence extraction means extracts one or more sentences as key sentences from the graph document,
The information processing method according to claim 1, wherein the summary generating means generates a summary that summarizes the input document based on the sentences extracted as key sentences by the key sentence extracting means.

Images