Document summarization device, method, and program

Abstract

To summarize a document more concisely. [Solution] The document summarization device according to the embodiment comprises an extraction unit, a classification unit, a determination unit, and a generation unit. The extraction unit performs natural language processing on the text contained in the document to extract a plurality of linguistic expressions and a first semantic relationship between the plurality of linguistic expressions from the text. The classification unit classifies the plurality of linguistic expressions into a plurality of clusters based on semantic similarity. The determination unit determines a second semantic relationship between the plurality of clusters based on the first semantic relationship. The generation unit generates a graph representing the plurality of clusters and the second semantic relationship.

Description

Technical Field

[0001] Embodiments of the present invention relate to a document summarization apparatus, method, and program.

Background Art

[0002] In the manufacturing industry, a document reporting troubles that occur during business (hereinafter referred to as a "trouble report") is customarily created. Generally, a trouble report records various events related to a trouble (e.g., phenomenon, investigation, cause, countermeasure, result) as text in natural language. The trouble report contributes to preventing recurrence of similar troubles and prompt resolution.

[0003] In a trouble report, the text related to various events can be enormous or complex. To summarize such a trouble report, the prior art performs natural language processing on the text included in the trouble report to extract a plurality of linguistic expressions and the semantic relationships between the plurality of linguistic expressions and presents them to the user.

[0004] However, a trouble report can include a large number of linguistic expressions. Therefore, when all the linguistic expressions and semantic relationships extracted from the trouble report are presented to the user, it is difficult for the user to understand the presented information at a glance.

Prior Art Documents

Patent Documents

[0005]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0006] The problem to be solved by the present invention is to summarize a document more concisely.

Means for Solving the Problems

[0007] The document summarization device according to this embodiment comprises an extraction unit, a classification unit, a determination unit, and a generation unit. The extraction unit performs natural language processing on the text contained in the document to extract a plurality of linguistic expressions and a first semantic relationship between the plurality of linguistic expressions from the text. The classification unit classifies the plurality of linguistic expressions into a plurality of clusters based on semantic similarity. The determination unit determines a second semantic relationship between the plurality of clusters based on the first semantic relationship. The generation unit generates a graph representing the plurality of clusters and the second semantic relationship. [Brief explanation of the drawing]

[0008] [Figure 1] Figure 1 is a functional configuration diagram of a document summarization device according to the first embodiment. [Figure 2] Figure 2 is a flowchart of the document summarization device according to the first embodiment. [Figure 3] Figure 3 shows a diagram illustrating natural language processing for text contained within a document. [Figure 4] Figure 4 shows the classification process for multiple linguistic expressions. [Figure 5] Figure 5 shows the process for determining semantic relationships between multiple clusters. [Figure 6] Figure 6 shows a graph. [Figure 7] Figure 7 is a functional configuration diagram of a document summarization device according to the second embodiment. [Figure 8] Figure 8 is a flowchart of a document summarization device according to the second embodiment. [Figure 9] Figure 9 shows the process of identifying a portion of the graph. [Figure 10] Figure 10 shows an example of a subgraph. [Figure 11] Figure 11 is a functional configuration diagram of a document summarization device according to the third embodiment. [Figure 12] Figure 12 is a flowchart of a document summarization device according to a third embodiment. [Figure 13]FIG. 13 is a diagram showing another example of a partial graph. [Figure 14] FIG. 14 is a functional configuration diagram of a document summarization apparatus according to a fourth embodiment. [Figure 15] FIG. 15 is a flowchart of a document summarization apparatus according to a fourth embodiment. [Figure 16] FIG. 16 is a diagram showing a conversion process from one language expression to another language expression. [Figure 17] FIG. 17 is a diagram showing a conversion process from a language expression to an image. [Figure 18] FIG. 18 is a diagram showing an example of a transition process of a display screen according to a modification. [Figure 19] FIG. 19 is a diagram showing another example of a transition process of a display screen according to a modification. [Figure 20] FIG. 20 is a hardware configuration diagram of a document summarization apparatus according to each embodiment.

Mode for Carrying Out the Invention

[0009] Hereinafter, each embodiment will be described with reference to the drawings. A plurality of parts to which the same reference numerals are assigned are regarded as being the same, and overlapping explanations will be omitted as appropriate.

[0010] (First Embodiment) FIG. 1 is a functional configuration diagram of a document summarization apparatus 1A according to a first embodiment. The document summarization apparatus 1A is an apparatus for summarizing a document. The document summarization apparatus 1A includes an acquisition unit 11, an extraction unit 12, a classification unit 13, a determination unit 14, and a generation unit 15.

[0011] The acquisition unit 11 is means for acquiring various data. The acquisition unit 11 acquires a document 2A from an external database or the like. The acquisition unit 11 transmits the document 2A to the extraction unit 12.

[0012] Document 2A is electronic document data (e.g., text data). Document 2A may be a trouble report. A trouble report records various events related to a trouble (e.g., phenomenon, investigation, cause, countermeasure, result) as text in natural language. A trouble report records a phenomenon as a trouble, an investigation to find the cause of the phenomenon, the cause found as a result of the investigation, a countermeasure against the cause, and a result caused by the countermeasure. A trouble report may be associated with a department, equipment number, manufacturing process, device, etc.

[0013] The extraction unit 12 is a means for extracting various data. The extraction unit 12 performs natural language processing on the text included in Document 2A received from the acquisition unit 11, and extracts a plurality of language expressions LE and a semantic relationship (hereinafter also referred to as "the first semantic relationship SR1") between the plurality of language expressions LE from the text. The extraction unit 12 may perform named entity extraction to extract a plurality of language expressions LE. The extraction unit 12 may perform relationship extraction to extract the first semantic relationship SR1. The extraction unit 12 transmits the extracted plurality of language expressions LE to the classification unit 13, and transmits the extracted first semantic relationship SR1 to the determination unit 14.

[0014] A language expression LE is a linguistic expression composed of a plurality of words. A language expression LE is a phrase, clause or sentence. A language expression LE is also referred to as "tag text".

[0015] The classification unit 13 is a means for classifying various data. The classification unit 13 classifies the plurality of language expressions LE received from the extraction unit 12 into a plurality of clusters CL according to semantic similarity. The classification unit 13 may perform hierarchical clustering or non-hierarchical clustering to classify the plurality of language expressions LE into a plurality of clusters CL. The classification unit 13 transmits the plurality of clusters CL to the determination unit 14 and the generation unit 15.

[0016] The determination unit 14 is a means for determining various types of data. Based on the first semantic relationship SR1 received from the extraction unit 12, the determination unit 14 determines the semantic relationships between multiple clusters CL received from the classification unit 13 (hereinafter also referred to as the "second semantic relationship SR2"). For the first cluster and the second cluster among the multiple clusters CL, the determination unit 14 determines the second semantic relationship SR2 between the first cluster and the second cluster based on the first semantic relationship SR1 between multiple first language expressions included in the first cluster and multiple second language expressions included in the second cluster. The determination unit 14 may also determine the first semantic relationship SR1 between one of the multiple first language expressions and one of the multiple second language expressions as the second semantic relationship SR2 between the first cluster and the second cluster. The determination unit 14 transmits the determined second semantic relationship SR2 to the generation unit 15.

[0017] The generation unit 15 is a means for generating various types of data. The generation unit 15 generates a graph 3A that represents multiple clusters CL received from the classification unit 13 and a second semantic relationship SR2 received from the determination unit 14. The generation unit 15 outputs the generated graph 3A to an external display device or the like.

[0018] Graph 3A is digitized graph data (e.g., image data). Graph 3A may represent clusters CL as nodes and the second semantic relationship SR2 between multiple clusters CL as edges. Graph 3A may be an undirected or directed graph.

[0019] Figure 2 is a flowchart of the document summarization device 1A according to the first embodiment. The document summarization device 1A performs the following steps S1A to S5A.

[0020] (Step S1A) First, the acquisition unit 11 acquires document 2A. For example, the acquisition unit 11 acquires document 2A from an external database or the like.

[0021] (Step S2A) Next, the extraction unit 12 performs natural language processing on the text contained in document 2A. For example, the extraction unit 12 performs named entity recognition and relational recognition on the text to extract multiple linguistic expressions LE and a first semantic relationship SR1 between the multiple linguistic expressions LE from the text (see Figure 3).

[0022] (Step S3A) Next, the classification unit 13 classifies the multiple linguistic expressions LE into multiple clusters CL. For example, the classification unit 13 performs hierarchical clustering or non-hierarchical clustering to classify the multiple linguistic expressions LE into multiple clusters CL based on semantic similarity (see Figure 4).

[0023] (Step S4A) Next, the determination unit 14 determines the semantic relationships (second semantic relationships SR2) between the multiple clusters CL. For example, the determination unit 14 determines the second semantic relationship SR2 between the multiple clusters CL based on the first semantic relationship SR1 between the multiple linguistic expressions LE (see Figure 5).

[0024] (Step S5A) Finally, the generation unit 15 generates graph 3A. For example, the generation unit 15 generates graph 3A based on multiple clusters CL and a second semantic relationship SR2 (see Figure 6).

[0025] Figure 3 shows the natural language processing performed on the text contained in document 2A. The extraction unit 12 performs (1) morphological analysis, (2) vectorization, (3) named entity recognition, and (4) relational extraction as natural language processing.

[0026] For example, the extraction unit 12 starts a series of processes on the source text, "Water leaked from the pipe, causing rust to form on the device, but there are no problems." (Stage ST21). First, the extraction unit 12 performs morphological analysis on the source text and divides it into multiple words (Stage ST22). Next, the extraction unit 12 vectorizes the divided words (Stage ST23). The extraction unit 12 may vectorize the words using a known language model (e.g., Word2Vec, BERT). As a result, the source text is vectorized into vectors V1 to V20. Each of the vectors V1 to V20 corresponds to a single word.

[0027] Next, the extraction unit 12 extracts language expressions LE from the vectorized words (stage ST24). The extraction unit 12 may also extract language expressions LE using a pre-trained language model (e.g., a neural network). As a result, three language expressions LE1, LE2, and LE3 are extracted from vectors V1 to V20. Language expression LE1 corresponds to vectors V1 to V6 and corresponds to the original text "Water leaked from the pipe". Language expression LE2 corresponds to vectors V9 to V15 and corresponds to the original text "Rust occurred on the device". Language expression LE3 corresponds to vectors V18 to V19 and corresponds to the original text "No problem".

[0028] Finally, the extraction unit 12 extracts a first semantic relationship SR1 between multiple linguistic expressions LE (stage ST25). The extraction unit 12 may extract the first semantic relationship SR1 using a pre-trained linguistic model (e.g., a neural network). The extraction unit 12 may also extract the first semantic relationship SR1 between two linguistic expressions LE based on the words present between those two linguistic expressions LE.

[0029] For example, the extraction unit 12 extracts a "causal relationship" between two linguistic expressions LE1 and LE2 based on the word "node" (corresponding to vector V7) that exists between the two linguistic expressions LE1 and LE2. This causal relationship may be directed from linguistic expression LE1 to linguistic expression LE2. Similarly, the extraction unit 12 extracts a "contradictory relationship" between two linguistic expressions LE2 and LE3 based on the word "mono no" (corresponding to vector V16) that exists between the two linguistic expressions LE2 and LE3.

[0030] Figure 4 shows the classification process for multiple linguistic expressions (LEs). The classification unit 13 performs (1) vectorization and (2) clustering as classification processes.

[0031] For example, the classification unit 13 starts a series of processes on the source text in the state shown in stage ST24 (see Figure 3) (stage ST31). First, the classification unit 13 vectorizes the language expressions LE extracted from the source text (stage ST32). The classification unit 13 may vectorize the language expressions LE using a known language model (e.g., Word2Vec, BERT). The classification unit 13 may vectorize the language expressions LE by averaging multiple vectors corresponding to multiple words in the language expressions LE. As a result, the three language expressions LE1, LE2, and LE3 are vectorized into three vectors VE1, VE2, and VE3, respectively.

[0032] Finally, the classification unit 13 clusters the multiple vectors VE, each corresponding to one of the multiple linguistic expressions LE (stage ST33). The classification unit 13 may cluster the multiple vectors VE that are adjacent to each other in the vector space using a hierarchical method (e.g., Ward's method) or a non-hierarchical method (e.g., K-means method).

[0033] The classification unit 13 may determine the accuracy of the clustering based on the results of clustering pairs of language expressions LE that should belong to the same (or different) clusters. For example, the classification unit 13 may quantitatively determine the accuracy of the clustering based on the distance between the clustered pairs of language expressions LE in the vector space. As a result, the three language expressions LE1, LE2, and LE3 are clustered into three clusters (cluster 1, cluster 2, and cluster 3).

[0034] The representation table ET shows the names of the three clusters and the multiple linguistic expressions LE classified into the three clusters. In addition to the three linguistic expressions LE1, LE2, and LE3, the representation table ET further shows the clustering results for other multiple linguistic expressions LE. Specifically, cluster 1 includes multiple linguistic expressions LE (e.g., water leak, water leaked from the pipe, there was a water leak from the pipe). Cluster 2 includes multiple linguistic expressions LE (e.g., rust, rust occurred on the device, it is rusty). Cluster 3 includes multiple linguistic expressions LE (e.g., no problem, no errors, no trouble).

[0035] A vector VE corresponding to a given linguistic expression LE indicates the meaning of that linguistic expression LE. Therefore, the classification unit 13 can collect multiple linguistic expressions LE with similar meanings into the same cluster by clustering multiple vectors VE. For example, cluster 1 includes multiple linguistic expressions LE with similar meanings to "water leak". Cluster 2 includes multiple linguistic expressions LE with similar meanings to "rusting". Cluster 3 includes multiple linguistic expressions LE with similar meanings to "no problem".

[0036] Figure 5 shows the process for determining the semantic relationships (second semantic relationship SR2) between multiple clusters CL. The determination unit 14 performs semantic relationship mapping as part of the determination process.

[0037] For example, the decision unit 14 starts processing using the source text in the state shown in stage ST25 (see Figure 3) and the expression table ET shown in stage ST33 (see Figure 4) (stage ST41). The decision unit 14 focuses on the three linguistic expressions LE1, LE2, and LE3, and the first semantic relationship SR1 between the three linguistic expressions LE1, LE2, and LE3. For example, the decision unit 14 notices that there is a "causal relationship" between linguistic expression LE1, which is contained in cluster 1, and linguistic expression LE2, which is contained in cluster 2. The decision unit 14 determines the "causal relationship" between linguistic expression LE1 and linguistic expression LE2 as the "causal relationship" between cluster 1 and cluster 2. Similarly, the decision unit 14 determines the "adverse relationship" between linguistic expression LE2 and linguistic expression LE3 as the "adverse relationship" between cluster 2 and cluster 3. In other words, the decision unit 14 maps the first semantic relationship SR1 between two linguistic expressions LE to the second semantic relationship SR2 between two clusters CL.

[0038] The relation table RT shows the second semantic relationship SR2 between the two clusters CL (stage ST42). When a semantic relationship exists from one cluster to another, that cluster is also called the "root cluster," and the other cluster is also called the "leaf cluster." The relation table RT has a 3x3 matrix formed by the combination of three root clusters and three leaf clusters. According to the relation table RT, there is a "causal relationship" between cluster 1 and cluster 2, and an "adverse relationship" between cluster 2 and cluster 3. The relation table RT may have similar information for other clusters different from clusters 1 to 3.

[0039] The determination unit 14 may also determine the strength of the semantic relationships between the first cluster (root cluster) and the second cluster (leaf cluster). For example, the determination unit 14 uses the number of first language expressions N(X) included in the first cluster, the number of second language expressions N(Y) included in the second cluster, and the number of first semantic relationships SR1 from the multiple first language expressions to the multiple second language expressions N(X→Y). The determination unit 14 may determine the strength of the semantic relationships between the first cluster and the second cluster using the formula "N(X→Y) / (N(X)×N(Y))". This formula calculates the proportion of the total number of semantic relationships N(X→Y) from the first cluster to the second cluster, relative to the total number of semantic relationships (N(X)×N(Y)) resulting from combinations between multiple first linguistic expressions and multiple second linguistic expressions.

[0040] Next, the determination unit 14 may determine whether the strength of the semantic relationship is greater than or equal to a threshold. If the strength is greater than or equal to the threshold, the determination unit 14 may determine the first semantic relationship SR1 from one of the plurality of first linguistic expressions to one of the plurality of second linguistic expressions as the second semantic relationship SR2 from the first cluster to the second cluster. The threshold may be set to a predetermined value, or it may be set to an arbitrary value by the user or the like.

[0041] Figure 6 shows Graph 3A. In Graph 3A, cluster CL in relation table RT is represented as node ND, and the second semantic relationship SR2 between multiple cluster CLs is represented as edge ED. Clusters 1 to 5 correspond to nodes ND1 to ND5, respectively. An edge ED from one cluster (root cluster) to another cluster (leaf cluster) is represented by arranging the number of the original cluster and the number of the other cluster in that order. For example, the edge ED from cluster 1 to cluster 2 is represented as "edge ED12".

[0042] A cluster may have multiple edge EDs as inputs. For example, cluster 2 may have two edge EDs, ED12 and ED42, as inputs. A cluster may have multiple edge EDs as outputs. For example, cluster 1 may have two edge EDs, ED12 and ED14, as outputs. Furthermore, there may be edge EDs that are output from a cluster and return to that cluster (i.e., self-loops). For example, edge ED22 is output from cluster 2 and returns to cluster 2.

[0043] According to the document summarization device 1A described above, the extraction unit 12 performs natural language processing on the text contained in document 2A to extract multiple linguistic expressions LE and a first semantic relationship SR1 between the multiple linguistic expressions LE from the text. The classification unit 13 classifies the multiple linguistic expressions LE into multiple clusters CL based on semantic similarity. The determination unit 14 extracts a second semantic relationship SR2 between the multiple clusters CL based on the first semantic relationship SR1. The generation unit 15 generates a graph 3A that represents the multiple clusters CL and the second semantic relationship SR2.

[0044] In other words, the document summarization device 1A classifies multiple linguistic expressions LE extracted from document 2A into multiple clusters CL, so that multiple linguistic expressions LE with similar meanings can be aggregated into the same cluster CL (or node ND). Furthermore, the document summarization device 1A aggregates the first semantic relationship SR1 between multiple linguistic expressions LE into a second semantic relationship SR2 (or edge ED) between multiple clusters CL, so that the number of first semantic relationships SR1 can be reduced to the number of second semantic relationships SR2. Therefore, the document summarization device 1A can summarize document 2A more concisely.

[0045] (Second embodiment) Figure 7 is a functional configuration diagram of the document summarization device 1B according to the second embodiment. In addition to the acquisition unit 11, extraction unit 12, classification unit 13, determination unit 14, and generation unit 15 of the document summarization device 1A, the document summarization device 1B further includes a storage unit 16 and a identification unit 17.

[0046] The acquisition unit 11 acquires another document 2B in addition to document 2A from an external database or the like. Another document 2B is the same as document 2A. The acquisition unit 11 transmits the acquired document 2A and another document 2B to the extraction unit 12.

[0047] The extraction unit 12 performs natural language processing on another text contained within another document 2B to extract from the other text a plurality of other linguistic expressions LEB and another first semantic relationship SRB1 between the plurality of other linguistic expressions LEB. The extraction unit 12 may also perform named entity recognition to extract the plurality of other linguistic expressions LEB. The extraction unit 12 may also perform relational extraction to extract the other first semantic relationship SRB1. The extraction unit 12 transmits the extracted plurality of linguistic expressions LEB and the other first semantic relationship SRB1 to the identification unit 17.

[0048] The generation unit 15 generates a subgraph 3B that represents the portion of graph 3A identified by the identification unit 17. The subgraph 3B is the same as graph 3A. The generation unit 15 outputs the generated subgraph 3B to an external display device or the like.

[0049] The storage unit 16 is a means for storing various types of data. The storage unit 16 stores the graph 3A generated by the generation unit 15. The storage unit 16 transmits the stored graph 3A to the generation unit 15 or the identification unit 17.

[0050] The identification unit 17 is a means for identifying various types of data. The identification unit 17 identifies portions from graph 3A that correspond to multiple other linguistic expressions LEB and another first semantic relation SRB1, and that represent multiple clusters CL and a second semantic relation SR2. The identification unit 17 transmits the identified portions to the generation unit 15.

[0051] Figure 8 is a flowchart of the document summarization device 1B according to the second embodiment. The document summarization device 1B may perform a series of processes similar to those of the document summarization device 1A. The document summarization device 1B performs the following steps S1B to S5B following step S5A (see Figure 2).

[0052] (Step S1B) First, the memory unit 16 stores graph 3A. For example, the memory unit 16 stores graph 3A generated in step S5A.

[0053] (Step S2B) Next, the acquisition unit 11 acquires another document 2B. For example, the acquisition unit 11 acquires another document 2B from an external database or the like.

[0054] (Step S3B) Next, the extraction unit 12 performs natural language processing on another text contained within another document 2B. For example, the extraction unit 12 performs named entity recognition and relational recognition on the other text to extract several other linguistic representations (LEBs) and another first semantic relation (SRB1) from the other text. Step S3B is the same as step S2A (see Figure 2).

[0055] (Step S4B) Next, the identification unit 17 identifies portions from graph 3A that correspond to other multiple linguistic expressions LEB and semantic relationships (another first semantic relationship SRB1). For example, the identification unit 17 identifies multiple linguistic expressions LE that correspond to other multiple linguistic expressions LEB from graph 3A, and identifies multiple clusters CL (or nodes ND) that encompass the identified multiple linguistic expressions LE. If a linguistic expression LE corresponding to a certain linguistic expression LEB does not exist in graph 3A, the identification unit 17 ignores that linguistic expression LEB. On the other hand, the identification unit 17 identifies edges ED that correspond to another first semantic relationship SRB1 between other multiple linguistic expressions LEB from graph 3A. If an edge ED corresponding to another first semantic relationship SRB1 does not exist in graph 3A, the identification unit 17 ignores that other first semantic relationship SRB1 (see Figure 9).

[0056] Subsequently, the identification unit 17 integrates the multiple clusters CL and edges ED identified from graph 3A to identify the parts corresponding to the other multiple linguistic expressions LEB and semantic relationships. Conversely, the identification unit 17 removes the remaining parts from graph 3A that do not correspond to those parts.

[0057] (Step S5B) Finally, the generation unit 15 generates the subgraph 3B. For example, the generation unit 15 generates the subgraph 3B based on the portion identified by the identification unit 17 (see Figure 10).

[0058] Figure 9 shows the process of identifying a portion from Graph 3A. For ease of explanation, the portion identified from Graph 3A is shown with thick lines and bold text. Each of clusters 1 to 5 in Graph 3A contains multiple linguistic expressions (LEs).

[0059] For example, the identification unit 17 identifies three linguistic expressions LE (water leak, rusty, no error) that correspond to multiple other linguistic expressions LEB, and identifies three clusters CL (cluster 1, cluster 2, cluster 3) that each contain these three linguistic expressions LE. On the other hand, the identification unit 17 identifies two edges ED (edge ​​ED12, edge ED23) that correspond to another first semantic relation SRB1. The identification unit 17 integrates the identified three clusters CL and two edges ED to identify a portion in graph 3A.

[0060] Figure 10 shows an example of subgraph 3B. Subgraph 3B represents a portion identified from graph 3A. This portion includes three clusters CL (cluster 1, cluster 2, cluster 3) and two edges ED (edges ED12, edge ED23). This portion may be highlighted in any manner (e.g., thick line, bold text, blinking).

[0061] According to the document summarization device 1B described above, the extraction unit 12 performs natural language processing on another text contained in another document 2B to extract from the other text a plurality of other linguistic expressions LEB and another first semantic relationship SRB1 between the plurality of other linguistic expressions LEB. The identification unit 17 identifies from the graph 3A a portion that corresponds to the plurality of other linguistic expressions LEB and another first semantic relationship SRB1, and that represents a plurality of clusters CL and a second semantic relationship SR2. The generation unit 15 generates a subgraph 3B that represents that portion.

[0062] Generally, another document 2B may contain a large number of linguistic expressions. Therefore, if all linguistic expressions and semantic relationships extracted from another document 2B are presented to the user, the user will have difficulty understanding the presented information.

[0063] The document summarization device 1B identifies portions of a pre-generated graph 3A that correspond to multiple linguistic expressions and semantic relationships extracted from another document 2B. The document summarization device 1B generates a subgraph 3B that represents the identified portions. In other words, the document summarization device 1B aggregates the multiple linguistic expressions and semantic relationships extracted from another document 2B into multiple clusters and semantic relationships in the subgraph 3B, thereby enabling a more concise summary of the other document 2B.

[0064] (Third embodiment) Figure 11 is a functional configuration diagram of the document summarization device 1C according to the third embodiment. Similar to the document summarization device 1B, the document summarization device 1C includes an acquisition unit 11, an extraction unit 12, a classification unit 13, a determination unit 14, a generation unit 15, a storage unit 16, and a identification unit 17.

[0065] The acquisition unit 11 acquires user information 2U in addition to document 2A and another document 2B from an external database or the like. The acquisition unit 11 transmits the acquired document 2A and another document 2B to the extraction unit 12 and transmits the acquired user information 2U to the identification unit 17.

[0066] User Information 2U is attribute information relating to the user for whom the subgraph 3B is presented. User Information 2U includes (1) personal information about the user, (2) skills information about the user's skills, and (3) work information about the user's work. Personal information includes name, age, gender, address, and contact information. Skills information includes work history, length of employment, and qualifications. Work information includes department, equipment number, manufacturing process, and equipment.

[0067] In particular, business information may include impact information regarding the degree of impact of problems caused by users during their work. Impact information includes frequency of occurrence, downtime, and damage.

[0068] The generation unit 15 generates a subgraph 3B. The generation unit 15 transmits the generated subgraph 3B to the storage unit 16 or the identification unit 17. The generation unit 15 generates a subgraph 3BT from the subgraph 3B that emphasizes the language expression LE identified by the identification unit 17. The generation unit 15 outputs the generated subgraph 3BT to an external display device or the like.

[0069] The storage unit 16 stores the subgraph 3B generated by the generation unit 15. The storage unit 16 transmits the stored subgraph 3B to the generation unit 15 or the identification unit 17.

[0070] The identification unit 17 identifies a language expression LE related to the user information 2U from multiple cluster CLs in the subgraph 3B based on the user information 2U. The identification unit 17 transmits the identified language expression LE to the generation unit 15.

[0071] Figure 12 is a flowchart of the document summarization device 1C according to the third embodiment. The document summarization device 1C may perform a series of processes similar to those of the document summarization device 1B. The document summarization device 1C performs the following steps S1C to S4C following step S5B (see Figure 8).

[0072] (Step S1C) First, the memory unit 16 stores the subgraph 3B. For example, the memory unit 16 stores the subgraph 3B generated in step S5B.

[0073] (Step S2C) Next, the acquisition unit 11 acquires user information 2U. For example, the acquisition unit 11 acquires user information 2U from an external database or the like.

[0074] (Step S3C) Next, the identification unit 17 identifies language expression LE related to user information 2U from the subgraph 3B. For example, the identification unit 17 identifies language expression LE related to user information 2U from multiple language expression LEs that are included in multiple cluster CLs in the subgraph 3B (see Figure 13).

[0075] (Step S4C) Finally, the generation unit 15 generates a subgraph 3BT that highlights the identified language expression LE. For example, the generation unit 15 highlights the identified language expression LE as a representative representation of the cluster CL. The generation unit 15 may also highlight the characters representing the language expression LE in any manner (e.g., bold, italics, blinking). On the other hand, the generation unit 15 may remove language expressions LE that were not identified from the cluster CL (see Figure 13).

[0076] Figure 13 shows another example of subgraph 3B (i.e., subgraph 3BT). For example, subgraph 3B is a summary about "Unit X," and it is assumed that the user to whom subgraph 3B is presented is in charge of "Unit Y." In this case, multiple cluster CLs in subgraph 3B contain linguistic expressions LE about "Unit X." On the other hand, multiple cluster CLs in subgraph 3B may also contain linguistic expressions LE about "Unit Y."

[0077] Therefore, the identification unit 17 identifies linguistic expressions LE related to "Unit Y" which the user is in charge of, from multiple cluster CLs in subgraph 3B. For example, the identification unit 17 identifies three linguistic expressions LE related to "Unit Y" (water leakage, rusting, no trouble) from three cluster CLs. The generation unit 15 generates a subgraph 3BT that enlarges and emphasizes the three identified linguistic expressions LE.

[0078] According to the document summarization device 1C described above, the identification unit 17 identifies language expressions LE related to the user information 2U from multiple clusters CL in the subgraph 3B based on the user information 2U relating to the user to whom the subgraph 3B is presented. The generation unit 15 generates a subgraph 3BT that highlights the identified language expressions LE.

[0079] The user views the subgraph 3BT on the display device. Firstly, if the user's "personal information" or "skills information" is used as user information 2U, the user can understand that a similar person caused the trouble and empathize with the trouble. Secondly, if the user's "work information" is used as user information 2U, the user can understand that a trouble related to their own work has occurred and empathize with the trouble. Thirdly, if the "impact information" of the trouble caused by the user is used as user information 2U, the user can understand how serious the presented trouble should be considered.

[0080] Generally, trouble reports are shared among multiple readers in a generalized or abstracted form. However, there is a concern that readers may not carefully read such trouble reports for reasons such as not finding them relevant to their own attributes or not being of interest to them.

[0081] Therefore, the document summarization device 1C uses user information 2U relating to the reader's (i.e., user's) attributes to identify and highlight linguistic expressions related to the user information 2U from the subgraph 3B, which is a summary of the trouble report. Consequently, the user is expected to carefully read the presented trouble report because they can perceive its relevance to or interest in their own attributes.

[0082] (Fourth embodiment) Figure 14 is a functional configuration diagram of the document summarization device 1D according to the fourth embodiment. In addition to the acquisition unit 11, extraction unit 12, classification unit 13, determination unit 14, generation unit 15, storage unit 16, and identification unit 17 provided by the document summarization device 1C, the document summarization device 1D further includes a conversion unit 18.

[0083] The conversion unit 18 is a means for converting various types of data into other types of data. The conversion unit 18 receives user information 2U from the acquisition unit 11 and a subgraph 3B from the generation unit 15. Firstly, the conversion unit 18 inputs the user information 2U and the language expressions LE included in multiple clusters CL in the subgraph 3B into a large-scale language model (or generation AI) and converts the input language expressions into language expressions LE related to the user information 2U. The conversion unit 18 outputs the subgraph 3B containing the converted language expressions LE (i.e., the converted graph 3C) to an external display device or the like.

[0084] Secondly, the conversion unit 18 inputs the language expressions LE, which are contained within multiple clusters CL in the subgraph 3B, into a large-scale language model (or generative AI) and converts the input language expressions into a 3D image. The conversion unit 18 outputs the 3D image to an external display device or the like.

[0085] Figure 15 is a flowchart of the document summarization device 1D according to the fourth embodiment. The document summarization device 1D may perform a series of processes similar to those of the document summarization device 1B. The document summarization device 1D performs the following steps S1D ​​to S4D following step S5B (see Figure 8).

[0086] (Step S1D) First, the memory unit 16 stores the subgraph 3B. For example, the memory unit 16 stores the subgraph 3B generated in step S5B. Step S1D is the same as step S1C (see Figure 12).

[0087] (Step S2D) Next, the acquisition unit 11 acquires user information 2U. For example, the acquisition unit 11 acquires user information 2U from an external database or the like. Step S2D is the same as step S2C (see Figure 12).

[0088] (Step S3D) Next, the conversion unit 18 converts the language expression LE in the subgraph 3B into a language expression LE (or image 3D) associated with the user information 2U. Firstly, the conversion unit 18 may input the user information 2U and the language expression LE to be converted into a large language model and convert the input language expression LE into a language expression LE associated with the user information 2U. At this time, the large language model may also be input a prompt instructing it to "convert to a language expression associated with user information" (see Figure 16).

[0089] Secondly, the conversion unit 18 may input the language expression LE to be converted into a large-scale language model and convert the input language expression LE into a 3D image. In this case, the large-scale language model may also be input a prompt instructing it to "convert to image" (see Figure 17).

[0090] (Step S4D) Finally, the conversion unit 18 outputs the converted graph 3C (or image 3D) to an external display device or the like. The conversion unit 18 may output both the converted graph 3C and the image 3D to the same display device or the like.

[0091] Figure 16 shows the process of converting one linguistic expression LE to another. For example, multiple cluster CLs in subgraph 3B do not include the linguistic expression LE related to "Machine Y" which is handled by the user to whom subgraph 3B is presented. In this case, the conversion unit 18 may convert the linguistic expression LE in each of the multiple cluster CLs into a fictitious linguistic expression LE related to a problem in "Machine Y" that has not actually occurred.

[0092] Firstly, the conversion unit 18 converts the language expression LE "Water leaked from the pipe" in cluster 1 to another language expression LE "Water leaked from pipe P of unit Y". Secondly, the conversion unit 18 converts the language expression LE "Rust occurred on the device" in cluster 2 to another language expression LE "Rust occurred on module M of unit Y". Thirdly, the conversion unit 18 converts the language expression LE "No problem" in cluster 3 to another language expression LE "No problem with module M of unit Y".

[0093] The converted graph 3C includes the language representation LE after conversion by the conversion unit 18. The converted language representation LE may be used as a representative notation for cluster CL.

[0094] Figure 17 shows the conversion process from linguistic expressions LE to 3D images. For example, the conversion unit 18 converts the linguistic expression LE "Water leaked from the pipe" in subgraph 3B into image 3D1 which represents this linguistic expression LE. Similarly, the conversion unit 18 converts the linguistic expression LE "No problem" in subgraph 3B into image 3D2 which represents this linguistic expression LE.

[0095] According to the document summarization device 1D described above, the conversion unit 18 inputs user information 2U relating to the user to whom the subgraph 3B is presented, and language expressions LE encompassed in multiple clusters CL in the subgraph 3B, into a large-scale language model, and converts the input language expressions LE into language expressions LE related to the user information 2U.

[0096] The user views the converted language expression LE (or converted graph 3C) on a display device. The converted language expression LE contains information relevant to the user. Therefore, the user is expected to carefully read the converted graph 3C, as they may find it relevant to or interesting to their own attributes.

[0097] Alternatively, according to the document summarization device 1D, the conversion unit 18 inputs language expressions LE, which are included in multiple clusters CL in the subgraph 3B, into a large-scale language model and converts the input language expressions LE into images 3D.

[0098] The 3D image visually represents the linguistic expression (LE). Therefore, users can more easily understand the content of the converted graph (3C) at a glance.

[0099] (modified version) According to the document summarization device 1B of the second embodiment, the generation unit 15 generates a subgraph 3B. At this time, the generation unit 15 may generate an interactive display screen that responds to user actions based on the causal relationships between multiple clusters CL in the subgraph 3B. The display screen may have choices in a gamebook or adventure game.

[0100] For example, the generation unit 15 focuses on multiple third clusters (root clusters) and multiple fourth clusters (leaf clusters) in the subgraph 3B. If a causal relationship exists between multiple third clusters and multiple fourth clusters, the generation unit 15 may generate a first display screen that includes linguistic expressions from each of the multiple third clusters. Furthermore, if one linguistic expression is selected on the first display screen, the generation unit 15 focuses on the fourth clusters (leaf clusters) that have a causal relationship with the third cluster (root cluster) that includes the selected linguistic expression. The generation unit 15 may generate a second display screen that includes linguistic expressions from the fourth clusters (see Figure 18).

[0101] On the other hand, the generation unit 15 may generate a first display screen that includes language expressions from a fifth cluster (root cluster) that does not have a causal relationship with multiple fourth clusters (leaf clusters). Furthermore, if a language expression from the fifth cluster is selected on the first display screen, the generation unit 15 does not need to generate a second display screen. Alternatively, the generation unit 15 may generate a display screen that indicates that it is not possible to transition to the next display screen (for example, a "game over screen") (see Figure 19).

[0102] Figure 18 shows an example of the transition process of the display screen in a modified example. The first display screen SC1A and the second display screen SC2 include a situation ST, a choice OP, and a cursor CR. Situation ST is a sentence that describes the current situation (or context). Choice OP are selectable actions (or commands) for situation ST. The cursor CR can be operated by the user through an input device.

[0103] In the first display screen SC1A, the situation ST includes a linguistic expression from a single cluster (root cluster) that has a causal relationship with multiple third clusters (leaf clusters). Specifically, the situation ST includes the sentence "Unit A is abnormal. What do you want to do?". Three options (OP) are presented as choices for this situation ST (replace unit B, update unit C, restart unit D). The user moves the cursor CR up and down to select the desired option (for example, "restart unit D"). In response to this selection, the first display screen SC1A transitions to the second display screen SC2.

[0104] In the second display screen SC2, the situation ST includes linguistic expressions from the fourth cluster (leaf cluster) that have a causal relationship with the third cluster (root cluster) that includes the selected option. Specifically, the situation ST includes the sentence "Unit D did not start. What will you do?". As options OP for this situation, several options (replace unit B, wait and see, ...) are presented. Options OP may also include linguistic expressions from the remaining third clusters that were not selected. The user moves the cursor CR up and down and selects the desired option (for example, "replace unit B"). In response to this selection, the second display screen SC2 transitions to another display screen as described above.

[0105] In this way, the generation unit 15 generates an interactive display screen that responds to user actions. This display screen is generated based on actual problems that have occurred in the past and the actions taken to address those problems. Therefore, users can virtually experience dealing with actual problems they have not experienced themselves, thereby improving their problem-solving skills. Furthermore, users can more easily understand the content of the problems at a glance and learn how to deal with problems in a fun, game-like way.

[0106] Figure 19 shows another example of the transition process of the display screen related to the modified example. In the first display screen SC1B, the situation ST includes the text "Unit A is malfunctioning. What do you want to do?". Four options OP are presented for this situation (replace unit B, update unit C, restart unit D, change the settings of unit E). The user moves the cursor CR up and down to select the desired option (for example, "change the settings of unit E").

[0107] However, the option "Change the settings of Unit E" is irrelevant or inappropriate as a solution to the problem "Anomaly in Unit A". In response to selecting this option (i.e., the dummy option), the first display screen SC1B does not transition to the second display screen SC2, but instead transitions to the third display screen SC3.

[0108] The third display screen SC3 contains the text "Game Over". Below this text are the Retry button B1 and the Exit button B2. The user moves the cursor CR left or right to select either the Retry button B1 or the Exit button B2. In response to the selection of the Retry button B1, the third display screen SC3 may transition to the first display screen SC1B. In response to the selection of the Exit button B2, the third display screen SC3 may be deleted.

[0109] In this way, the generation unit 15 introduces irrelevant or inappropriate options into the first display screen SC1B. If this option is selected, the first display screen SC1B transitions to a third display screen SC3 in which there are no options for dealing with the problem. Therefore, the user can easily understand that the selected option is irrelevant or inappropriate for dealing with the problem.

[0110] Figure 20 is a hardware configuration diagram of the document summarization device 1 according to each embodiment. The document summarization device 1 comprises a CPU 101, RAM 102, ROM 103, storage 104, display device 105, input device 106, and communication device 107 as its components. Each component is connected to each other via an internal bus so as to be able to communicate with each other. The document summarization device 1 may include at least some of each component.

[0111] The CPU 101 is a processor that executes various processes according to a program. The CPU 101 uses a predetermined area of ​​the RAM 102 as a working area. The CPU 101 realizes each part (e.g., acquisition unit 11, extraction unit 12, classification unit 13, determination unit 14, generation unit 15, identification unit 17, conversion unit 18) by reading and executing each program stored in the ROM 103 or storage 104. Each part may be realized by dedicated hardware circuits (e.g., ASIC, PLD, FPGA). Each part may be realized on-premise or in the cloud. The CPU 101 is an example of a processing unit.

[0112] RAM102 is a memory that stores various types of data in a rewritable format. For example, RAM102 is SDRAM (Synchronous Dynamic Random Access Memory). RAM102 is an example of a storage unit 16.

[0113] ROM103 is a memory that stores various types of data in a non-rewritable format. ROM113 is an example of a storage unit 16.

[0114] The storage 104 is various types of storage media. The storage 104 may also be a drive device that writes various types of data to the storage media or reads various types of data from the storage media. The storage 104 may be controlled by the CPU 101. The storage 104 is an example of the storage unit 16.

[0115] The display device 105 is a device that displays various types of data. The display device 105 may be an LCD (Liquid Crystal Display). The display device 105 displays various types of data based on display signals from the CPU 101. The display device 105 is an example of a display unit.

[0116] Input device 106 is a device that receives various input operations from the user. Input device 106 may be a mouse or a keyboard. Input device 106 receives the operations entered by the user as instruction signals and transmits the instruction signals to the CPU 101. Input device 106 is an example of an input unit.

[0117] The communication device 107 communicates with external devices via a network in accordance with the control of the CPU 101. The communication device 107 is an example of a communication unit.

[0118] While several embodiments of the present invention have been described, these embodiments are presented as examples only and are not intended to limit the scope of the invention. These novel embodiments can be carried out in a variety of other forms, and various omissions, substitutions, and modifications can be made without departing from the spirit of the invention. These embodiments and their variations are included in the scope and spirit of the invention, as well as in the claims of the invention and its equivalents. [Explanation of Symbols]

[0119] 1,1A,1B,1C,1D…Document summarization device, 2A…Document, 2B…Another document, 2U…User information, 3A…Graph, 3B,3BT…Partial graph, 3C…Converted graph, 3D,3D1,3D2…Image, 11…Acquisition unit, 12…Extraction unit, 13…Classification unit, 14…Decision unit, 15…Generation unit, 16…Storage unit, 17…Specification unit, 18…Conversion unit, 101…CPU, 102…RAM, 103…ROM, 104…Storage, 105…Display device, 106…Input device, 107…Communication device, B…Button, CL…Cluster, CR…Cursor, ED…Edge, ET…Representation table, LE…Language representation, ND…Node, OP…Choice, RT…Relationship table, SC…Display screen, SR…Semantic relationship, ST…Situation, V…Vector

Claims

1. An extraction unit that performs natural language processing on the text contained in a document to extract multiple linguistic expressions and a first semantic relationship between the multiple linguistic expressions from the text, A classification unit that classifies the aforementioned multiple linguistic expressions into multiple clusters based on semantic similarity, A determination unit that determines a second semantic relationship between the plurality of clusters based on the first semantic relationship, A generation unit that generates a graph representing the plurality of clusters and the second semantic relationship, A document summarization device equipped with the following features.

2. The determination unit determines, with respect to the first cluster and the second cluster among the plurality of clusters, the second semantic relationship between the first cluster and the second cluster based on the first semantic relationship between the plurality of first linguistic expressions included in the first cluster and the plurality of second linguistic expressions included in the second cluster. The document summarization device according to claim 1.

3. The determination unit determines the first semantic relationship between one of the plurality of first linguistic expressions and one of the plurality of second linguistic expressions as the second semantic relationship between the first cluster and the second cluster. The document summarization device according to claim 2.

4. The determination unit determines the strength of the second semantic relationship from the first cluster to the second cluster using the number of the plurality of first linguistic expressions, the number of the plurality of second linguistic expressions, and the number of the first semantic relationships from the plurality of first linguistic expressions to the plurality of second linguistic expressions. The document summarization device according to claim 2.

5. The determination unit determines, if the strength is greater than or equal to a threshold, the first semantic relationship from one of the plurality of first linguistic expressions to one of the plurality of second linguistic expressions as the second semantic relationship from the first cluster to the second cluster. The document summarization device according to claim 4.

6. It further comprises a specific part, The extraction unit performs the natural language processing on another text contained within another document to extract from the other text a plurality of other linguistic expressions and a first semantic relationship between the plurality of other linguistic expressions. The specified part identifies from the graph a portion that corresponds to the other multiple linguistic expressions and the other first semantic relationship, and that represents the multiple clusters and the second semantic relationship. The generation unit generates a subgraph that represents the portion. The document summarization device according to claim 1.

7. The identifying unit identifies linguistic expressions related to the user information from the multiple clusters in the subgraph based on the user information relating to the user to whom the subgraph is presented. The generation unit generates the subgraph that emphasizes the identified linguistic expression. The document summarization device according to claim 6.

8. The system further comprises a conversion unit that inputs user information relating to the user on whom the subgraph is presented and language expressions included in the multiple clusters in the subgraph into a large-scale language model, and converts the input language expressions into language expressions related to the user information. The document summarization device according to claim 6.

9. The user information includes at least one of the following: personal information relating to the user; skills information relating to the user's skills; and business information relating to the user's work. The document summarization device according to claim 7 or claim 8.

10. The system further comprises a conversion unit that inputs language expressions included in the multiple clusters in the subgraph into a large-scale language model and converts the input language expressions into images. The document summarization device according to claim 6.

11. The generation unit generates a first display screen that includes linguistic expressions from each of the multiple third clusters if a causal relationship exists between the multiple third clusters and the multiple fourth clusters in the subgraph. When one language expression is selected on the first display screen, a second display screen is generated that includes language expressions from a fourth cluster having a causal relationship with the third cluster that includes the selected language expression. The document summarization device according to claim 6.

12. The generation unit generates the first display screen which includes language expressions from a fifth cluster that does not have a causal relationship with the plurality of fourth clusters, If the language expression from the fifth cluster is selected on the first display screen, the second display screen is not generated. The document summarization device according to claim 11.

13. Computers Natural language processing is performed on the text contained in the document to extract multiple linguistic expressions and a first semantic relationship between the multiple linguistic expressions from the text. The aforementioned multiple linguistic expressions are classified into multiple clusters based on their semantic similarity, Based on the first semantic relationship, a second semantic relationship is determined between the multiple clusters. A graph is generated that represents the plurality of clusters and the second semantic relationship. Document summarization methods.

14. On the computer, An extraction function that performs natural language processing on text contained in a document to extract multiple linguistic expressions and a first semantic relationship between the multiple linguistic expressions from the text, A classification function that classifies the aforementioned multiple linguistic expressions into multiple clusters based on semantic similarity, A decision function that determines a second semantic relationship between the plurality of clusters based on the first semantic relationship, A generation function that generates a graph representing the plurality of clusters and the second semantic relationship, A document summarization program that achieves this.

Images