Information processing device, information processing program, and information processing method

Abstract

An information processing device, an information processing program, and an information processing method are provided that process documents while taking into consideration their logical structure. [Solution] The present invention relates to an information processing device, an information processing program, and an information processing method for summarizing a document, and the information processing device of the present invention is characterized by comprising: a graph embedding generation means for acquiring graph embeddings expressed with one or more embeddings based on a graph document in a graph format; and a summary generation unit for generating a summary of the graph document based on the graph embeddings.

Description

[Technical Field]

[0001] The present invention relates to an information processing device, an information processing program, and an information processing method, and can be applied to, for example, a process for summarizing a document. [Background technology]

[0002] Methods for automatically summarizing documents (data containing sentences) can be broadly divided into two types: "extraction-based summarization" and "abstraction-based summarization." Extraction-based summarization is a method of extracting important sentences from a document and combining them to create a summary. Abstraction-based summarization is a method of generating a summary by inputting the document to be summarized into a generative model such as a neural network. Extraction-based summarization involves extracting some sentences from the document and linking them together, which can result in summaries with unnatural connections between sentences. On the other hand, abstraction-based summarization can generate more concise and natural summaries by using words that are not contained in the document to be summarized. Furthermore, with the recent rise of large-scale language models, the performance of abstraction-based summarization has further improved.

[0003] A conventional extractive summarization technique is the method described in Patent Document 1. The method described in Patent Document 1 solves the problems of extractive summarization by dividing the text to be summarized into blocks based on topics and using abstract summarization for each block. Furthermore, the method described in Patent Document 1 structures the summary results of each block to create summaries that are easy for users to read. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Japanese Patent Application Publication No. 2023-34235 Summary of the Invention [Problem to be solved by the invention]

[0005] However, the method described in Patent Document 1 has a problem in that it is only able to list the summary results of different blocks for each topic, and is unable to create a summary that summarizes the contents of the entire document.

[0006] Furthermore, the method described in Patent Document 1 has the problem that it is not possible to create a summary that takes into consideration the importance of the topic or the logical relationships between blocks.

[0007] In view of the above problems, there is a demand for an information processing device, an information processing program, and an information processing method that process documents while taking into consideration the logical structure of the documents. [Means for solving the problem]

[0008] The information processing device of the first invention is characterized by having a graph embedding generation means for acquiring graph embeddings expressed in one or more embeddings based on a graph document in graph format, and a summary generation unit for generating a summary of the graph document based on the graph embeddings.

[0009] In the first aspect of the present invention, the graph embedding generation means may perform a feature quantification process to acquire features of the nodes and edges that make up the graph document, generate node embeddings for the nodes that make up the graph document after the feature quantification process, and acquire the graph embedding based on the node embeddings. Also, in the first aspect of the present invention, the graph embedding generation means may, in the feature quantification process, generate sentence embeddings for the text of the nodes that make up the graph document and acquire them as features of the nodes, and generate word embeddings for the text of the edges that make up the graph document and acquire them as features of the edges. Furthermore, in the first aspect of the present invention, the graph embedding generation means may aggregate the created node embeddings to generate a single embedding and acquire it as the graph embedding.

[0010] The second information processing program of the present invention is characterized in that it causes a computer to function as a graph embedding generation means that obtains graph embeddings expressed in one or more embeddings based on a graph document in graph format, and a summary generation unit that generates a summary of the graph document based on the graph embeddings.

[0011] The third aspect of the present invention is an information processing method performed by an information processing device, characterized in that the information processing device comprises a graph embedding generation means and a summary generation unit, the graph embedding generation means obtains graph embeddings expressed in one or more embeddings based on a graph document in graph format, and the summary generation unit generates a summary of the graph document based on the graph embeddings. [Effects of the Invention]

[0012] According to the present invention, it is possible to provide an information processing device, an information processing program, and an information processing method that process documents while taking into consideration their logical structure. [Brief explanation of the drawings]

[0013] [Figure 1] 1 is a block diagram showing a functional configuration of a document summarization device according to an embodiment. [Figure 2] FIG. 2 is a diagram illustrating a structure of a graph document according to the embodiment. [Figure 3] 1 is a flowchart showing the overall operation of a document summarization device according to an embodiment. [Figure 4] 10 is a flowchart illustrating the operation of a pre-processing unit according to the embodiment. [Figure 5] 10 is a flowchart illustrating the operation of a graph embedding generation unit according to the embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0014] (A) First embodiment An embodiment of an information processing device, an information processing program, and an information processing method according to the present invention will be described below in detail with reference to the drawings. In this embodiment, an example in which the information processing device, the information processing program, and the information processing method according to the present invention are applied to a document summarization device will be described.

[0015] (A-1) Configuration of the main embodiment FIG. 1 is a block diagram showing the functional configuration of a document summarizing device 10 according to this embodiment.

[0016] The document summarization device 10 is a device that performs document summarization processing to summarize input data including documents composed of multiple sentences, and outputs output data including the summaries. Hereinafter, the documents included in the input data will be referred to as "input documents." Note that a summary includes one or more sentences.

[0017] The document summarization device 10 includes a document input unit 100, a preprocessing unit 200, a graph embedding generation unit 300, a summary generation unit 400, and a summary output unit 500.

[0018] The document summarization device 10 may be configured entirely from hardware (for example, 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.

[0019] The document input unit 11 performs a process of accepting input of input data (data including an input document). The means for receiving the input data in the document input unit 11 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).

[0020] The preprocessing unit 200 has a function of preprocessing (details of which will be described later) input documents included in the input data. The preprocessing unit 200 has a document format determination unit 210 and a graph generation unit 220.

[0021] The document format determination unit 210 performs a process of determining the format of an input document (hereinafter referred to as "document format determination process"). Specifically, as the document format determination process, the document format determination unit 210 performs a process of determining whether the input document corresponds to a graph document (a document having structural information in a graph format / graph structure) or plain text having no particular structural information.

[0022] The graph generation unit 220 performs a conversion process to convert an input document that is not in graph format (for example, if it is plain text) into a document in graph format (i.e., a graph document). The graph generation unit 220 has a graph generation unit 220 that generates nodes from the input document, and a node generation unit 221 that generates edges that indicate discourse relationships between the generated nodes.

[0023] Here, the structure of the graph document in this embodiment will be explained with reference to FIG.

[0024] FIG. 2 is a diagram showing the structure of a graph document in this embodiment.

[0025] Figure 2 shows an example of the structure of a graph document based on the plain text input document "It will rain tomorrow. The field trip will be canceled. The probability of precipitation is 90%."

[0026] 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 (relationship) between the nodes at both ends (between divided sentences). In FIG. 2, each node is represented by a rectangular symbol, and the edges between nodes are illustrated by arrows. In the graph document of this embodiment, each edge is assigned a directionality (orientation) in 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.

[0027] 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 in various graph formats (middleware that configures databases) can be applied.

[0028] In the example graph document structure shown in Figure 2, three segmented 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 segmented sentence. Also, in the example graph document structure shown in Figure 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 graph document structure shown in Figure 2, an edge E2 labeled "Additional Information" (logical relationship) is set from node X1 (It will rain tomorrow) to node X3 (The probability of precipitation is 90%).

[0029] As described above, in this embodiment, the graph document has a structure as shown in the example of FIG.

[0030] The graph embedding generation unit 300 performs a process (hereinafter referred to as "graph embedding process") to generate data expressed using one or more embeddings (hereinafter simply referred to as "graph embedding") for the graph document output from the preprocessing unit 200, and outputs the generated graph embedding.

[0031] The summary generation unit 400 processes the graph-embedded data output by the graph-embedding generation unit 300 to generate a summary (a summary for the input document) (hereinafter referred to as "summary generation processing").

[0032] The summary output unit 500 performs processing to output data of the summary generated by the summary generation unit 400. The summary output unit 500 also outputs data including at least the summary as output data. The method by which the summary output unit 500 outputs the output data is not limited. For example, the summary output unit 500 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 on a display device such as a monitor, or may print the output from a printing device such as a printer.

[0033] (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.

[0034] FIG. 3 is a flowchart showing the overall operation of the document summarizing device 10.

[0035] Here, it is assumed that input data including an input document is first input to the document summarization device 10 (S100). In the document summarization device 10, the input document is acquired by the document input unit 100. The document input unit 100 supplies the input document included in the input data to the preprocessing unit 200.

[0036] Next, the preprocessing unit 200 preprocesses the supplied input document, holds a graph document (graph-formatted input document), and supplies it to the graph embedding generation unit 300 (S200). The operation of the preprocessing unit 200 will be described in detail later.

[0037] Next, the graph embedding generation unit 300 generates a graph embedding based on the supplied graph document and supplies it to the summary generation unit 400 (S300). The operation of the graph embedding generation unit 300 will be described in detail later.

[0038] Next, the summary generation unit 400 generates a summary (a summary sentence for the input document) based on the supplied graph embeddings and supplies the summary to the summary output unit 500 (S400). The process by which the summary generation unit 400 generates a summary from graph embeddings is not limited. For example, the summary generation unit 400 may generate a summary by inputting the graph embeddings into pre-trained large language models (LLMs).

[0039] Next, the summary output unit 500 performs processing to output the supplied summary in a predetermined format (S500).

[0040] The document summarizing device 10 performs the above-described processing as a whole.

[0041] Next, the details of the operation of the preprocessing unit 200 will be described with reference to FIG.

[0042] FIG. 3 is a flowchart showing the details of the operation of the preprocessing unit 200.

[0043] In the preprocessing unit 200, an input document is first supplied to the document format determination unit 210. Then, the document format determination unit 210 performs a document format determination process to determine the document format of the input document (S201). Specifically, the document format determination unit 210 determines whether the input document is a graph document (a document having a predetermined graph format) or plain text (simple character string data without structural information). The method by which the document format determination unit 210 performs the document determination process is not limited, and various methods can be applied. The document format determination unit 210 may perform the document determination process based on information such as the file format of the input document (for example, information about the file extension and data format).

[0044] If the input document is determined to be plain text in step S201, the preprocessing unit 200 supplies the plain text to the node generation unit 221 of the graph generation unit 220. The node generation unit 221 performs a process (hereinafter referred to as a "document segmentation process") of segmenting the supplied plain text into predetermined units (e.g., units each having a semantically coherent meaning) (hereinafter, these segmented units are referred to as "segmented sentences") and generates a set of nodes having each segmented sentence (text) as a value (S202). The method by which the node generation unit 221 segments the input document is not limited, and various methods can be used. The node generation unit 221 may extract segmented sentences according to a predetermined algorithm (e.g., extracting segmented sentences based on symbols such as punctuation marks or line break codes included in the input document), or may extract segmented sentences using a machine learning model that has previously trained pairs of "plain text" and "collections of segmented sentences" (collections of correct segmented sentences) as training data. Furthermore, for example, the node generating unit 221 may perform document segmentation processing by dividing an input document into discourse units using a discourse relation analyzer that has undergone machine learning.

[0045] Next, the edge generation unit 222 generates edges indicating discourse relations between nodes for the set of nodes generated by the node generation unit 221 (S203). The method by which the edge generation unit 222 generates edges is not limited, and various methods can be used. The edge generation unit 222 can be realized, for example, by using a machine learning model (e.g., a machine-learned discourse relation analyzer, etc.) that has been trained in advance using training data, as in the case of node generation. Note that in this embodiment, the process of generating a graph document from an input document is described as being divided into a node generation step and an edge generation step, but node generation and edge generation may also be performed in a single step (e.g., nodes and edges may be generated simultaneously using a machine-learned discourse relation analyzer).

[0046] As described above, the preprocessing unit 200 (graph generating unit 220) can hold a graph document corresponding to the plain text as the input document, based on the set of nodes generated in step S202 and the set of edges generated in step S203. Furthermore, if the input document itself is a graph document, the preprocessing unit 200 may hold the graph document without performing any additional processing.

[0047] Then, the preprocessing unit 200 outputs (supplies to the graph embedding generation unit 300) the graph document that has been held (S204).

[0048] Next, the detailed processing of the graph embedding generation unit 300 will be described.

[0049] FIG. 5 is a flowchart showing the detailed processing of the graph embedding generation unit 300.

[0050] First, the graph embedding generation unit 300 generates features (e.g., codes / vectors / tokens corresponding to the divided sentences) of each node (the text of the divided sentences set in each node) that constitutes the graph document (S301). At this time, the specific method by which the graph embedding generation unit 300 acquires the features of each node is not limited. The graph embedding generation unit 300 may be realized by, for example, generating sentence embeddings from the text of each node using various language models (e.g., embedding layers of language models; the applicable language models are not limited).

[0051] Next, the graph embedding generation unit 300 generates features (e.g., codes / vectors / tokens corresponding to words) of each edge (words as labels set for each edge) that constitute the graph document (S302). At this time, the specific method by which the graph embedding generation unit 300 acquires the features of each edge is not limited. The graph embedding generation unit 300 may be realized by, for example, generating word embeddings from the text (word text) of each edge using various language models (e.g., embedding layers of language models; the suitable language model is not limited).

[0052] Through the above processing, the graph embedding generation unit 300 can obtain a graph document in which the text (text of divided sentences or words) of each element (each node and each edge) is characterized (encoded / vectorized / tokenized).

[0053] Next, the graph embedding generation unit 300 generates node embeddings for each node of the graph document based on the graph document in which each element has been characterized (S303). The node embeddings are obtained by encoding each node in graph structure data, such as a graph document, into a vector representation (a vector representation reflecting the graph structure). The specific method by which the graph embedding generation unit 300 obtains the node embeddings is not limited. The graph embedding generation unit 300 may use, for example, various trainable graph neural networks (hereinafter simply referred to as "GNNs"). Examples of GNNs that the graph embedding generation unit 300 uses to obtain the node embeddings include PyTorch Geometric (https: / / pytorch-geometric.readthedocs.io / en / latest / index.html) and Deep Graph Library (https: / / docs.dgl.ai / ).

[0054] Next, the graph embedding generator 300 aggregates the generated node embeddings into one embedding, generating a graph embedding (one embedding corresponding to the graph document) (S304). At this time, the specific method by which the graph embedding generator 300 aggregates the node embeddings of each node and expresses them as one embedding (graph embedding) is not limited. For example, the graph embedding generator 300 may aggregate the node embeddings of each node by taking the average of the node embeddings of each node using a method such as average pooling.

[0055] The graph embedding output by the graph embedding generator 300 may be a collection of multiple node embeddings. In other words, the graph embedding is data made up of one or more embeddings.

[0056] (A-3) Effects of the embodiment According to this embodiment, the following effects can be achieved.

[0057] In this embodiment, the document summarization device 10 acquires graph embeddings (data representing a graph document using one or more embeddings) for an input document and processes the acquired graph embeddings to create a summary. The graph embeddings for an input document (graph document) can be said to be codes that clearly reflect the logical structure of the input document. Therefore, by performing a process of creating a summary using graph embeddings, the document summarization device 10 has the effect of being able to generate a summary that takes into account the logical structure of the input document.

[0058] Furthermore, in the document summarization device 10 of this embodiment, the graph embedding generation unit 300 generates graph embeddings based on a graph document, and the summary generation unit 400 generates a summary based on the graph embeddings. These steps can be processed using different learning models (e.g., language models, GNNs, etc.). Therefore, in this case, the language model used in the summary generation unit 400 is fixed, and only the language model and GNN used in the graph embedding generation unit 300 can be updated, such as through additional learning. As mentioned above, it is desirable to use a large-scale language model (LLM) for the summary generation unit 400, but updating a large-scale language model is very costly. On the other hand, the language model and GNN used in the graph embedding generation unit 300 can be constructed and updated through a much smaller learning process (e.g., supervised learning in which a graph document is paired with a correct summary) than the large-scale language model (LLM). In other words, when summarizing using a single large-scale language model, as in conventional summarization processing, it is necessary to update the large-scale language model itself to optimize it for summarization processing. However, in the document summarization device 10 of this embodiment, the large-scale language model (LLM) applied to the summary generation unit 400 is fixed, and the entire system can be updated simply by updating the learning model of the graph embedding generation unit 300 (i.e., the entire system can be updated at a lower cost than conventional methods).

[0059] (B) Other embodiments The present invention is not limited to the above-described embodiments, and may include modified embodiments such as those exemplified below.

[0060] (B-1) In the above embodiment, the summary generation unit 400 of the document summarization device 10 has been described as an example of performing abstract summarization using a large-scale language model for graph embedding. However, this is not limited to this example, and for example, extraction summarization using a classifier configured with a neural network may also be used.

[0061] (B-2) In the above embodiment, if the input document (input data) input to the document summarization device 10 is only a graph document, the preprocessing unit 200 may be excluded. [Explanation of symbols]

[0062] 10... document summarization device, 100... document input unit, 200... preprocessing unit, 210... document format determination unit, 220... graph generation unit, 221... node generation unit, 222... edge generation unit, 300... graph embedding generation unit, 400... summary generation unit, 500... summary output unit

Claims

1. a graph embedding generator for obtaining a graph embedding expressed by one or more embeddings based on a graph document in a graph format; a summary generator for generating a summary of the graph document based on the graph embedding; An information processing device comprising:

2. 2. The information processing device according to claim 1, wherein the graph embedding generation means performs a feature quantification process to acquire features of nodes and edges that constitute the graph document, generates node embeddings for the nodes that constitute the graph document after the feature quantification process, and acquires the graph embedding based on the node embeddings.

3. 3. The information processing device according to claim 2, wherein the graph embedding generation means, in the feature quantification process, generates sentence embeddings for the text of the nodes that make up the graph document and acquires them as features of the nodes, and generates word embeddings for the text of the edges that make up the graph document and acquires them as features of the edges.

4. 3. The information processing apparatus according to claim 2, wherein the graph embedding generating means aggregates the created node embeddings to generate one embedding and acquires the single embedding as the graph embedding.

5. Computer, a graph embedding generator for obtaining a graph embedding expressed by one or more embeddings based on a graph document in a graph format; a summary generator for generating a summary of the graph document based on the graph embedding; and a processor for processing the information.

6. An information processing method performed by an information processing device, the information processing device comprises a graph embedding generation unit and a summary generation unit; The graph embedding generating means obtains a graph embedding expressed by one or more embeddings based on a graph document in a graph format; The summary generator generates a summary of the graph document based on the graph embedding.

1. An information processing method comprising:

Images