JP7625201B2 - Knowledge model creation support device - Google Patents

Description

The present invention relates to a knowledge model creation support device.

Patent Document 1 describes a system for constructing knowledge models that describe knowledge held by experts in various fields, such as engineering, medicine, pharmacy, agriculture, and biology. The knowledge model uses key terms in the field as factors and expresses the terms in the field and their relationships by interconnecting the factors in a network form. Patent Document 2 describes a knowledge model in the field of machining.

JP 2018-147351 A JP 2020-049606 A

When creating a knowledge model, workers need to extract terms that define factors and then understand the relationships between those terms. However, extracting terms and understanding the relationships between the extracted terms requires manual setup, particularly by an experienced person, so creating a knowledge model is not easy.

The present invention aims to provide a knowledge model creation support device that can easily create knowledge models.

One aspect of the present invention is an apparatus for supporting the creation of a knowledge model that is configured with a plurality of factors defined by terms and relationship information between the factors, the apparatus comprising:
a term-relationship database including a plurality of terms and relationships of the plurality of terms;
a term-relationship database creation unit that creates the term-relationship database by extracting the terms and the relationships from document data using a term-relationship extraction model;
a knowledge model creation unit that creates the knowledge model based on the term-relationship database,
The term-relationship extraction model is
a document data acquisition unit for acquiring the document data;
a term extraction unit that extracts terms from the document data;
a trigger word storage unit for storing a trigger word that leads to a relationship between two terms;
a pair extraction unit that creates term pair information from the terms extracted by the term extraction unit;
a relationship extraction unit that extracts a relationship between the two terms extracted by the term extraction unit based on the trigger word and the pair information when the trigger word is included in the document data ,
the document data includes a nested structure including a plurality of terms;
The nested structure is a structure that is composed of a plurality of inclusive terms and a nested outer term that combines the plurality of inclusive terms,
the term extraction unit extracts, when the document data includes the nested structure, the inclusive term and the nested outer term in the nested structure from the document data;
when the document data includes a nested structure, the pair extraction unit creates the pair information by using only the nested outer term in the nested structure without using the inclusive term in the nested structure;
The relationship extraction unit is in a knowledge model creation support device that, when the document data includes the nested structure, treats the nested outer term in the nested structure as a relationship extraction target and does not treat the inclusive term as a relationship extraction target.

Another aspect of the present invention is an apparatus for supporting creation of a knowledge model that is configured by a plurality of factors defined by terms and relationship information between the factors, the apparatus comprising:
a term-relationship database including a plurality of terms and relationships of the plurality of terms;
a term-relationship database creation unit that creates the term-relationship database by extracting the terms and the relationships from document data using a term-relationship extraction model;
a knowledge model creation unit that creates the knowledge model based on the term-relationship database,
The term-relationship extraction model is
a document data acquisition unit for acquiring the document data;
a feature representation generating unit that generates a feature representation in which a feature of the document data is expressed as a vector based on each token constituting the document data acquired by the document data acquiring unit;
a term extraction unit that extracts terms from the document data;
a trigger word storage unit for storing a trigger word that leads to a relationship between two terms;
a relationship extraction unit that extracts a relationship between two terms extracted by the term extraction unit based on the trigger word when the trigger word is included in the document data,
The term extraction unit and the relationship extraction unit are in a knowledge model creation support device that extracts the terms and the relationships by sharing the characteristic expressions generated by the characteristic expression generation unit.

According to the above knowledge model creation support device, a term-relationship database is created using a term-relationship extraction model that extracts terms and relationships between terms from document data. First, the term-relationship extraction model automatically extracts terms from document data. Furthermore, the term-relationship extraction model pre-stores trigger words that lead to relationships between two terms, and when the document data contains the trigger words, extracts the relationship between the two terms based on the trigger words. In this way, the term-relationship extraction model can automatically extract terms from document data, and can automatically extract relationships between terms contained in the document data by taking into account the preset trigger words. Therefore, the term-relationship database can be created without manual work. Then, once the term-relationship database has been created, a knowledge model can be created by referring to the term-relationship database, and therefore the knowledge model can be easily created.

A knowledge network diagram is shown. 1 is a functional block diagram showing an overall configuration of a knowledge model creation support device; FIG. 2 is a diagram showing a first database (term-relationship database) constituting the knowledge model creation support device. 4 is a diagram illustrating relationship labels relating to relationships in the first database shown in FIG. 3. 1 is a functional block diagram showing a term-relationship extraction model in a first database creation unit constituting a knowledge model creation support device. FIG. FIG. 13 is a diagram for explaining a nested structure applied in the term extraction unit. FIG. 4 is a diagram showing trigger words stored in a trigger word storage unit. FIG. 13 is a diagram showing a detailed configuration of a term-relationship extraction model. FIG. 1 is a functional block diagram of a term-relationship extraction model in a first learning phase. FIG. 13 is a functional block diagram of a term-relationship extraction model in the second learning phase. 11 is a diagram showing a drawing GUI window showing an application example of a knowledge model creation unit constituting the knowledge model creation support device; FIG. FIG. 13 is a diagram showing a drawing GUI window illustrating another application example of the knowledge model creation unit. FIG. 13 is a diagram showing a drawing GUI window illustrating another application example of the knowledge model creation unit. FIG. 13 is a diagram showing a drawing GUI window illustrating another application example of the knowledge model creation unit.

(1. Overview of the knowledge model)
A knowledge model describes and stores knowledge related to information in a given field in a specified format. That is, a knowledge model is mainly composed of multiple factors defined by the terminology of the field and information on the relationships between the factors. For example, the knowledge model may be knowledge related to a technical field.

A knowledge model is represented, for example, by data containing information about factors and data containing relationship information. Conceptually, a knowledge model expresses the relationships between multiple factors by connecting them to each other in the form of a network. In other words, a knowledge model stores knowledge (including know-how) in various technical fields as explicit knowledge, and can also be updated.

Technical fields include engineering, medicine, pharmacy, agriculture, biology, and other fields. In particular, technical fields include the field of machining, which is included in the field of engineering. Here, the field of machining includes, for example, cutting and grinding. Furthermore, it is particularly useful for the knowledge model to describe knowledge about technical information held by experts in each technical field.

For example, in the field of machining, an operator determines the cutting speed, depth of cut, and other machining conditions by taking into account various information such as the material of the workpiece, the material of the tool, the machining accuracy, and the machining cycle time. In this case, the knowledge model is a model of the operator's thought process when the operator determines the cutting speed, depth of cut, and other machining conditions by inputting various information such as the material of the workpiece, the material of the tool, the machining accuracy, and the machining cycle time.

In other words, in addition to the workpiece material, tool material, machining requirements (machining accuracy, machining cycle time, etc.), cutting speed, cutting depth, etc., the knowledge model defines each of the technical elements that appear in the thought process as factors, and defines the relationships between the factors.

For example, the knowledge model is used as follows: When an operator inputs the necessary information about the workpiece material, tool material, and machining requirements (machining accuracy, machining cycle time, etc.) as input factors in the knowledge model, information about the cutting speed, cutting depth, etc. as output factors is output.

(2. Example of knowledge network diagram 100)
As described above, the knowledge model is conceptually expressed in the form of a network. An example of a knowledge network diagram 100 in which a knowledge model is expressed in the form of a network diagram will be described with reference to Fig. 1. In this example, a knowledge network diagram 100 relating to a knowledge model in the field of machining will be taken as an example.

As shown in FIG. 1, the knowledge network diagram 100 comprises a plurality of node figures 110 and link figures 120 that connect the node figures 110 to each other. The node figures 110 are represented by any figure such as a box, a figure including text, an icon, etc. The node figures 110 represent factors in the knowledge model. The link figures 120 are represented by straight lines, curves, hook lines, etc. In this example, the link figures 120 are represented by arrow lines to specify the directionality of the relationships. The link figures 120 represent relationships that connect factors in the knowledge model. In the knowledge network diagram 100 shown in FIG. 1, the node figures 110 are all represented by boxes in which text can be written, and the link figures 120 are represented by arrow lines.

Here, factors are defined by technical terms. In addition, multiple factors may have technical inclusion relationships (also called hierarchical relationships, parent-child relationships, or master-slave relationships) or may have technical heterogeneous relationships. In other words, the relationships between factors are classified into the two types mentioned above.

For example, the workpiece specifications include the thermal properties of the workpiece, hardness of the workpiece, elongation of the workpiece, etc. In other words, as factors having a technical inclusion relationship, the workpiece specifications are the higher-level conceptual factors, and the thermal properties of the workpiece, hardness of the workpiece, elongation of the workpiece, etc. are the lower-level conceptual factors. For example, factors having a technically heterogeneous relationship include the thermal properties of the workpiece and the required heat resistance of the tool. In what follows, the relationship between two factors having a technical inclusion relationship will be referred to simply as an inclusion relationship, and the relationship between two factors having a technically heterogeneous relationship will be referred to simply as a heterogeneous relationship.

Then, for the link graphic 120, the first link graphic 121, which represents an inclusive relationship, and the second link graphic 122, which represents a heterogeneous relationship, are displayed separately. In other words, the first link graphic 121 and the second link graphic 122 are displayed in different ways.

In FIG. 1, the first link graphic 121, which indicates an inclusion relationship, is represented by a frame line indicating the area of the superordinate conceptual factor, and the subordinate conceptual factor is arranged within the frame line representing the first link graphic 121. Note that the first link graphic 121 may also be represented by an L-shape connecting node graphics 110 arranged adjacently above and below and slightly shifted to the left and right, in addition to a frame line. In this case, of the two node graphics 110 connected by the first link graphic 121, the node graphic 110 located at the top corresponds to the superordinate conceptual factor.

In addition, in FIG. 1, the second link graphic 122 representing the heterogeneous relationship is represented by a straight line, a broken line, or the like, connecting the node graphics 110 placed at any position (up, down, left, right) away from each other. The second link graphic 122 is represented by an arrow line indicating the directionality of the definition between the factors.

(3. Configuration of knowledge model creation support device 1)
The knowledge model creation support device 1 is a support device for creating the above-mentioned knowledge model. The configuration of the knowledge model creation support device 1 will be described with reference to FIG.

The knowledge model creation support device 1 includes a first DB creation unit 2, a first DB 3, a second DB creation unit 4, a second DB 5, a third DB creation unit 6, a third DB 7, and a knowledge model creation unit 8. DB is an abbreviation for database. The knowledge model creation support device 1 is configured to include three DB creation units 2, 4, 6 and three DBs 3, 5, 7, but it may also be one of each, two of each, or four or more of each.

The first DB creation unit 2 creates the first DB 3, which is a term-relationship DB, by extracting multiple terms and relationships between the terms from document data (text data) using a term-relationship extraction model described below. In other words, when document data (text data) is input to the first DB creation unit 2, the term-relationship extraction model is executed, and multiple terms and relationships between the terms are automatically extracted. The first DB 3 stores the terms and relationships created by the first DB creation unit 2. The terms here refer to information that can be used to create factors in the above-mentioned knowledge model, and the relationships refer to information that can be used to create relationship information between factors in the above-mentioned knowledge model.

The second DB creation unit 4 is configured using a model, such as word2vec, that is different from the term-relationship extraction model that configures the first DB creation unit 2. The second DB creation unit 4 creates the second DB 5 by extracting relationships between multiple terms from document data using word2vec. The second DB 5 stores the relationships created by the second DB creation unit 4.

The third DB creation unit 6 is configured by a model different from the models configuring the first DB creation unit 2 and the second DB creation unit 4, for example, a known knowledge graph. The third DB creation unit 6 creates the third DB 7 by extracting relationships between multiple terms from the document data using the knowledge graph. The third DB 7 stores the relationships created by the third DB creation unit 6.

The knowledge model creation unit 8 creates a knowledge model using the terms and the relationships between the terms stored in the first DB 3, and further using the relationships between the terms stored in each of the second DB 5 and the third DB 7 as auxiliary data. For example, the knowledge model creation unit 8 sets multiple terms stored in the first DB 3 as factor candidates. Then, the knowledge model creation unit 8 uses each DB 3, 5, and 7 to provide candidates for locations to place the factors to be registered, provide factor candidates that have a relationship with the factor of interest, and verify knowledge models that have already been created. The knowledge model creation unit 8 can perform some of its work automatically, and other parts manually. However, it is also possible for the knowledge model creation unit 8 to fully automate the creation of knowledge models.

(4. Example of first DB3)
An example of the first DB (term-relationship DB) will be described with reference to Figures 3 and 4. As shown in Figure 3, the first DB 3 stores a reference term A, a related term B having a relationship with the reference term A, and a relationship label indicating the relationship between the reference term A and the related term B. For example, W1 as the reference term A has relationships with W2, W4, and W10, and the respective relationships (relationship labels) are Positive, Negative, and Positive.

As shown in FIG. 4, four types of relationship labels are defined: Positive, Negative, Sub, and Relation. Positive represents a relationship in which the larger the reference term A, the larger the related term B. Negative represents a relationship in which the larger the reference term A, the smaller the related term B. Sub represents a relationship in which the reference term A is a type of related term B. Relation represents a qualitative relationship between the reference term A and the related term B. Note that the relationship labels are not limited to the above four types, and may include other types and can be freely set.

(5. Configuration of First DB Creation Unit 2)
The configuration of the first DB creation unit 2 will be described with reference to Figures 5 to 7. The first DB creation unit 2 is configured by a term-relationship extraction model. As shown in Figure 5, the first DB creation unit 2 includes a document data acquisition unit 11, a characteristic expression generation unit 12, a term extraction unit 13, a term output unit 14, a trigger word storage unit 15, a pair extraction unit 16, a relationship extraction unit 17, and a relationship output unit 18.

The document data acquisition unit 11 acquires document data as text data. The characteristic representation generation unit 12 generates a characteristic representation that expresses the characteristics of the document data as vectors, based on each token that constitutes the document data acquired by the document data acquisition unit 11. The term extraction unit 13 extracts terms from the document data. In detail, the term extraction unit 13 extracts terms included in the document data using the characteristic representation generated by the characteristic representation generation unit 12. The term output unit 14 outputs the terms extracted by the term extraction unit 13.

Here, the term extraction unit 13 extracts terms taking into consideration the nesting structure 20. The nesting structure will be described with reference to FIG. 6. The nesting structure 20 is a structure in which a term contains another term. The nesting structure 20 is composed of multiple contained terms 22, 23 and a nested outer term 21 that combines multiple contained terms 22, 23.

For example, the machining term "cutting" contains the machining terms "cutting" and "processing". In this case, "cutting" is the nested outer term 21, and "cutting" and "processing" are the contained terms 22, 23. The nested outer term 21 and the contained terms 22, 23 that make up the nested structure have, for example, a superior-subordinate relationship, an attribute relationship, a subject-predicate relationship, etc.

As an example of a sentence from the acquired document data, "When the cutting speed increases, the cutting temperature increases." will be described. In this example sentence, the term extraction unit 13 extracts the terms "cutting speed," "cutting," "speed," "increase," "cutting temperature," "temperature," and "increase." In other words, "cutting speed" and "cutting temperature" are extracted as nested outer terms 21, and "cutting," "speed," and "temperature" are extracted as contained terms 22 and 23.

The trigger word storage unit 15 stores preset trigger words. A trigger word is a keyword that derives a relationship between two terms. A trigger word is, for example, a keyword that indicates a change in a term that represents a physical quantity. For example, as shown in FIG. 7, the trigger word storage unit 15 stores "increase," "decrease," "increase," "decrease," "rise," "fall," "contain," etc.

As an example of a sentence from the acquired document data, we will explain "When the cutting speed increases, the cutting temperature increases." In this example sentence, the trigger words are "increase" and "increase." The trigger words are keywords that can derive a relationship (positive relationship label) between the two terms "cutting speed" and "cutting temperature" in which if one increases, the other increases (see Figure 4).

The pair extraction unit 16 creates pairs from the terms extracted by the term extraction unit 13, and formats the data to be used by the relationship extraction unit 17 described below. When constructing a nested structure 20, the term extraction unit 13 extracts a nested outer term 21 and inclusive terms 22, 23, but the pair extraction unit 16 uses only the largest nested outer term 21 and does not use the inclusive terms 22, 23. When the number of terms to be paired is n, the pair extraction unit 16 creates _n C ₂ pairs.

When the document data contains a trigger word, the relationship extraction unit 17 extracts the relationship between the two terms extracted by the term extraction unit 13 based on the trigger word. When extracting the relationship, the relationship extraction unit 17 uses the trigger word stored in the trigger word storage unit 15 and the term pair information created by the pair extraction unit 16. Therefore, when the document data contains a nested structure 20 (shown in FIG. 6), the relationship extraction unit 17 extracts the relationship from the nested outer term 21 in the nested structure 20, but does not extract the relationship from the inclusive terms 22 and 23.

Furthermore, the relationship extraction unit 17 extracts the relationship between two terms using the characteristic representation generated by the characteristic representation generation unit 12. In other words, the above-mentioned term extraction unit 13 and the relationship extraction unit 17 share the characteristic representation generated by the characteristic representation generation unit 12 to extract terms and relationships. The relationship output unit 18 outputs the relationship between the terms extracted by the relationship extraction unit 17.

(6. Detailed configuration of term-relationship extraction model)
A detailed configuration of the term-relationship extraction model constituting the first DB creation unit 2 will be described with reference to Fig. 8. Fig. 8 shows the feature expression generation unit 12, the term extraction unit 13, the pair extraction unit 16, and the relationship extraction unit 17. An explanation will be given using an example sentence, "When the cutting speed increases, the cutting temperature increases,..."

(6-1. Feature Representation Generation Unit 12)
The feature representation generating unit 12 constitutes a shared unit for the term extracting unit 13 and the relationship extracting unit 17. The feature representation generating unit 12 divides the acquired document data into tokens and acquires a token string x. For example, each of "cut", "shave", and "speed" is one token. Next, a token representation H1 is acquired from the token string x. The token representation H1 is expressed by a string of vectors corresponding to the token string x. To acquire the token representation H1, a pre-training model for the token string x is used. For example, the token representation H1 is acquired as shown in formula (1) by using BERT (Bidirectional Encoder Representations from Transformers), which is one of the pre-training models.

Then, the acquired token representation H1 is used as an input to a convolutional neural network (CNN) to acquire an intermediate representation H2 as shown in formula (2). In this example, the intermediate representation H2 corresponds to the feature representation generated by the feature representation generation unit 12.

The intermediate representation H2 is generated using the set parameters. Here, the intermediate representation H2 is a representation used for both term extraction and relationship extraction. Therefore, as will be described later, in the learning phase, learning is performed based on the term extraction loss generated in the term extraction by the term extraction unit 13 and the relationship extraction loss generated in the relationship extraction by the relationship extraction unit 17. In other words, the parameters for generating the intermediate representation H2 are updated using the losses generated in each of the term extraction and relationship extraction.

(6-2.Term extraction unit 13)
The term extraction unit 13 focuses on the difference between the nested outer term 21 and the inclusive terms 22 and 23 that make up the nested structure 20 being the number of tokens that make up the term, and extracts the number of tokens (the length of the token) ) is a model that outputs two values for each input.

First, the intermediate representation H2 is used as input, and a convolution operation is performed to obtain a representation of a token string of length i, as shown in formula (3). In formula (3), spanCNNi represents a convolution operation using a filter with kernel size i, and _nentity corresponds to the number of tokens that make up a term and is a hyperparameter given before learning. The obtained intermediate representation Vi is an intermediate representation of _nentities corresponding to the number of tokens 1, 2, 3, ..., n.

Next, as shown in equation (4), a common fully connected layer is applied to each of the n _entity intermediate representations Vi corresponding to the number of tokens to generate an intermediate representation Li.

The output of the full combination is a sequence of token length corresponding to each kernel size of n _entities . As shown in FIG. 8, the contents of the output are such that a 1 is set at the start position of the token that constitutes the term of the kernel size. For example, the term "cutting speed" is composed of three tokens, and the start position of "cutting speed" is the position of the token "cut", so the token length is "3" and a 1 (black circle) is set at the token position of "cut". In addition, the term "speed" is composed of one token, and the start position is the token "speed", so the token length is "1" and a 1 (black circle) is set at the token position of "speed".

(6-3. Pair Extraction Unit 16)
As described above, the pair extraction unit 16 creates pairs of terms extracted by the term extraction unit 13. However, when the extracted terms form a nested structure 20, the pair extraction unit 16 extracts only the largest nested external term 21 as a relationship extraction target. For example, as shown in FIG. 8, "cutting speed", "cutting", and "speed" are extracted by the term extraction unit 13, and in this case, only "cutting speed" is extracted as a relationship extraction target. Therefore, in FIG. 8, the token number length is "3" and 1 is set at the position of the token "cut", and the positions corresponding to "cutting" and "speed" corresponding to the inclusive terms 22 and 23 are set to 0. The same is true for "cutting temperature". In other words, here, a pair of "cutting speed" and "cutting temperature" is created. When n terms are extracted from one sentence of document data, two terms are selected to create a pair, so _n C ₂ pairs are created.

Next, a relational label (see FIG. 4) is assigned to each pair created by referring to the annotation file, and a triplet of "relational label, term A, term B" is created. Furthermore, the pair extraction unit 16 also creates a term position vector PEi for term A and term B. The term position vector is a vector of the number of tokens with a length of one sentence, with 1 at the position where a term exists and 0 at the position where it is not a term. In FIG. 8, for one term of the pair, "cutting speed," 1 is placed at the positions of "cut," "cut," and "speed," and 0 is placed at the remaining positions. For the other term of the pair, "cutting temperature," 1 is placed at the positions of "cut," "cut," and "temperature," and 0 is placed at the remaining positions. In this way, the term position vector PEi is created by the pair extraction unit 16.

(6-4. Relationship Extraction Unit 17)
The relationship extraction unit 17 is a model that takes trigger words into consideration as described above. First, trigger word information is added to the intermediate representation H2 generated by the feature representation generation unit 12 as a sharing unit, and is used as an input to the Multi Head Attention layer. First, Vtrig is generated by referring to the trigger word storage unit 15 (shown in FIG. 5). Vtrig is a vector whose length is the token length (SeqLen) in one sentence and whose dimension is 1. For the corresponding token, 1 is defined if the token is a trigger word, and 0 is defined if the token is not a trigger word. Then, as shown in formula (5), an embedding process is performed on Vtrig using a matrix initialized with weights based on a normal distribution with a mean of 0 and a variance of 1.

Next, as shown in equation (6), the intermediate representation H2 is concat- ed against Htrig in the dimension direction, and input to the fully connected layer to generate Q (query).

K (key) is generated according to formula (7), and V (value) is generated according to formula (8).

Then, the generated Q, K, and V are input to a multi-head attention layer to generate intermediate representations H3 and W. In this attention layer, Q (query) is the search source (target), K (key) is the search destination (source), and V (value) is the score.

Here, Multi Head Attention adds trigger word information to Q (query). Since information is added to the query, the trigger word is added as "something whose relevance within the sentence we want to know." Also, considering that K (key) and V (value) contain information from the intermediate representation H2, the resulting intermediate representation H3 can be interpreted as "an expression that pays strong attention to tokens related to the trigger word among the tokens in the entire sentence."

Next, as shown in equations (10) and (11), the intermediate representation H3 is passed through two skip connections and two convolutional layers to generate intermediate representation H4.

Next, as shown in formula (12), the term position vector PEi and intermediate representation H4 created by the pair extraction unit 16 are used to extract expressions E1 and E2 of two terms to be used for relationship extraction. Here, PEi represents the i-th term position vector, and * represents the element product.

Next, as shown in equation (13), Maxpooling is performed on the obtained E1 and E2, and each is shaped into a single vector. Here, E1' is a term composed of multiple tokens.

Next, as shown in equation (14), E1' is input to the fully connected layer to generate a term representation E1".

Next, as shown in equation (15), the vectors of E1" and E2" are summed, and the probability of each relationship that exists between the terms is output through the fully connected layer as shown in equation (16). In other words, Lrel is a vector whose length is the number of types of relationships, the probability of the i-th type of relationship is represented by the value of the i-th element of Lrel, and the relationship with the highest probability in the Lrel vector represents the relationship between the two terms.

(7. Learning Phase)
(7-1. Overview)
The above-mentioned term-relationship extraction model needs to acquire a large amount of document data and perform machine learning. In particular, in this example, in the term-relationship extraction model, parameters in the feature representation generation unit 12 shared by the term extraction unit 13 and the relationship extraction unit 17 are learned.

(7-2. First learning phase)
The first learning phase as the first learning will be described with reference to Fig. 9. In the first learning phase, the term extraction loss Lency generated in the term extraction by the term extraction unit 13 is used to update the generation parameters of the intermediate representation H2 in the feature representation generation unit 12, thereby performing the first learning. That is, in the first learning phase, learning is performed by a single-task learning process that updates the generation parameters based on the term extraction loss Lency. In Fig. 9, functions that are not used in the first learning phase are indicated by dashed lines.

Here, for example, cross entropy loss is used as the term extraction loss Lentity. In this case, the term extraction loss Lentity is expressed by equation (17). n represents the number of model outputs, i.e., the number of filters prepared, and l represents the number of tokens in one sentence. The difference between the predicted value of l for each token and the correct answer y is defined as cross entropy. The term extraction loss Lentity is then calculated by adding up the outputs corresponding to each filter. Note that, in addition to cross entropy loss, the term extraction loss Lentity can also use losses obtained by other similar loss calculation formulas.

(7-3. Second learning phase)
Following the first learning phase, the second learning phase as the second learning will be described with reference to Fig. 10. In the second learning phase, a term extraction loss Lentity generated in the extraction of terms by the term extraction unit 13 and a relationship extraction loss Lrelation generated in the extraction of relationships by the relationship extraction unit 17 are used. In the second learning phase, the second learning is performed by updating the generation parameters of the intermediate representation H2 in the feature representation generation unit 12 based on the term extraction loss Lentity and the relationship extraction loss Lrelation. That is, in the second learning phase, learning is performed by a multi-task learning process that updates the generation parameters based on the term extraction loss Lentity and the relationship extraction loss Lrelation.

Here, the relationship extraction loss Lrelation is expressed by equation (18). r is the number of types of relation labels (see Figure 4). Since a loss occurs for each relation, the sum of all of these is the relationship extraction loss Lrelation. Furthermore, Nentity is the number of constituent terms in the nested structure 20 targeted for term extraction.

Then, using the term extraction loss Lentity expressed in the above formula (17) and the relationship extraction loss Lrelation shown in formula (18), the overall loss Lall is generated as shown in formula (19).

Here, to express the loss, for example, cross entropy loss can be used, as in the first learning phase. Cross entropy loss is expressed by equation (20). Class represents the label that is the target of relationship extraction.

(8. Example of Processing by Knowledge Model Creation Unit 8)
Next, examples of processing by the knowledge model creation unit 8 will be described with reference to Figs. 11 to 14. However, the processing by the knowledge model creation unit 8 is not limited to the following, and various processing is possible. Furthermore, the knowledge model creation unit 8 can be fully automated.

(8-1, first example)
The processing of the first example will be described with reference to Fig. 11. The knowledge model creation unit 8 draws a knowledge network diagram composed of node figures representing factors and link figures representing relationships between the factors in a drawable drawing GUI window 30. The drawing operation can be performed by the operator.

When an operator draws a knowledge network diagram in the drawing GUI window 30, the operator first determines a factor of interest. The knowledge model creation unit 8 then displays in the drawing GUI window 30 placement candidates for node figures of terms that have a relationship with the factor of interest, based on the information on the relationships in DBs 3, 5, and 7. At this time, multiple placement candidates can also be displayed. In particular, by using multiple DBs 3, 5, and 7, it becomes possible to output placement candidates from multiple perspectives. Then, by determining the placement candidates, the operator can draw (place) the factor of interest in the knowledge network diagram in the drawing GUI window 30.

(8-2. Second Example)
The processing of the second example will be described with reference to Fig. 12. In the drawing GUI window 30, the operator selects an already drawn factor of interest. The knowledge model creation unit 8 then displays terms having a relationship with the factor of interest as factor candidates based on the information on the relationships in the DBs 3, 5, and 7. The operator then selects from among the factor candidates, and the selected factor candidate is placed in the drawing GUI window 30 in a state associated with the factor of interest.

(8-3. Third Example)
The processing of the third example will be described with reference to Fig. 13. The worker inputs document data into the DB creation units 2, 4, and 6. The knowledge model creation unit 8 then displays a plurality of terms obtained based on the first DB 3 (term-relationship DB) in the GUI window 30. Then, when the worker selects one of the displayed plurality of terms, the knowledge model creation unit 8 displays, in the GUI window 30, terms that have a relationship with the selected term based on the information on the relationships in the DBs 3, 5, and 7.

In other words, the term selected by the worker and terms related to that term are displayed in the GUI window 30. Then, when the worker selects multiple displayed terms, the knowledge model creation unit 8 draws a knowledge network diagram in the GUI window 30.

(8-4. Fourth Example)
The processing of the fourth example will be described with reference to Fig. 14. An operator inputs document data into the DB creation units 2, 4, and 6. Then, the knowledge model creation unit 8 displays existing similar knowledge model candidates in the GUI window 30. When there are multiple similar knowledge model candidates, the similar knowledge model candidate selected by the operator is displayed in the GUI window 30.

Then, the knowledge model creation unit 8 displays, in the GUI window 30, a number of terms obtained based on the first DB3 (term-relationship DB) based on the input document data. Then, when the operator selects from the displayed number of terms, the knowledge model creation unit 8 displays, in the GUI window 30, terms that have a relationship to the selected term based on the information on the relationships in DB3, 5, and 7.

In other words, the GUI window 30 displays the similar knowledge model, as well as the term selected by the operator and terms related to that term. At this time, when the term selected by the operator is set as a factor of interest, the knowledge model creation unit 8 displays link figures that have no relationship to the factor of interest in the similar knowledge model displayed in the GUI window 30, for example, by using an X mark. Furthermore, when the term selected by the operator is not included in the similar knowledge model, the knowledge model creation unit 8 may newly draw the selected term.

(9. Effects)
According to the knowledge model creation support device 1, DB3 is created using a term-relationship extraction model that extracts terms and relationships between terms from document data. First, the term-relationship extraction model automatically extracts terms from document data. Furthermore, the term-relationship extraction model pre-stores trigger words that lead to relationships between two terms, and when the trigger words are included in the document data, extracts the relationship between the two terms based on the trigger words. In this way, the term-relationship extraction model can automatically extract terms from document data, and can automatically extract relationships between terms included in the document data by taking into account the preset trigger words. Therefore, DB3 can be created without manual work. Then, if DB3 can be created, a knowledge model can be created by referring to DB3, and therefore a knowledge model can be easily created.

1: Knowledge model creation support device, 2: First database creation unit, 3: First database, 4: Second database creation unit, 5: Second database, 6: Second database creation unit, 7: Third database, 8: Knowledge model creation unit, 11: Document data acquisition unit, 12: Feature expression generation unit, 13: Term extraction unit, 14: Term output unit, 15: Trigger word storage unit, 16: Pair extraction unit, 17: Relationship extraction unit, 18: Relationship output unit, 20: Nested structure, 21: Nested outer term, 22, 23: Inclusive term, 100: Knowledge network diagram, 110: Node diagram, 120: Link diagram, 121: First link diagram, 122: Second link diagram, Lentity: Term extraction loss, Lrelation: Relationship extraction loss

Claims (8)

An apparatus for supporting the creation of a knowledge model composed of a plurality of factors defined by terms and relationship information between the factors, comprising:
a term-relationship database including a plurality of terms and relationships of the plurality of terms;
a term-relationship database creation unit that creates the term-relationship database by extracting the terms and the relationships from document data using a term-relationship extraction model;
a knowledge model creation unit that creates the knowledge model based on the term-relationship database,
The term-relationship extraction model is
a document data acquisition unit for acquiring the document data;
a term extraction unit that extracts terms from the document data;
a trigger word storage unit for storing a trigger word that leads to a relationship between two terms;
a pair extraction unit that creates term pair information from the terms extracted by the term extraction unit;
a relationship extraction unit that extracts a relationship between the two terms extracted by the term extraction unit based on the trigger word and the pair information when the trigger word is included in the document data ,
the document data includes a nested structure including a plurality of terms;
The nested structure is a structure that is composed of a plurality of inclusive terms and a nested outer term that combines the plurality of inclusive terms,
the term extraction unit extracts, when the document data includes the nested structure, the inclusive term and the nested outer term in the nested structure from the document data;
when the document data includes a nested structure, the pair extraction unit creates the pair information by using only the nested outer term in the nested structure without using the inclusive term in the nested structure;
A knowledge model creation support device, wherein when the document data includes the nested structure, the relationship extraction unit treats the nested outer term in the nested structure as a relationship extraction target, and does not treat the contained term as a relationship extraction target . The knowledge model creation support device according to claim 1, wherein the trigger word is a keyword that represents a change in a term that represents a physical quantity. The term-relationship extraction model further comprises:
a feature representation generating unit that generates a feature representation in which a feature of the document data is expressed by a vector based on each token constituting the document data acquired by the document data acquiring unit,
3 . The knowledge model creation support device according to claim 1 , wherein the term extraction unit and the relationship extraction unit extract the terms and the relationships by sharing the characteristic expressions generated by the characteristic expression generation unit. An apparatus for supporting the creation of a knowledge model composed of a plurality of factors defined by terms and relationship information between the factors, comprising:
a term-relationship database including a plurality of terms and relationships of the plurality of terms;
a term-relationship database creation unit that creates the term-relationship database by extracting the terms and the relationships from document data using a term-relationship extraction model;
a knowledge model creation unit that creates the knowledge model based on the term-relationship database,
The term-relationship extraction model is
a document data acquisition unit for acquiring the document data;
a feature representation generating unit that generates a feature representation in which a feature of the document data is expressed as a vector based on each token constituting the document data acquired by the document data acquiring unit;
a term extraction unit that extracts terms from the document data;
a trigger word storage unit for storing a trigger word that leads to a relationship between two terms;
a relationship extraction unit that extracts a relationship between two terms extracted by the term extraction unit based on the trigger word when the trigger word is included in the document data,
A knowledge model creation support device, wherein the term extraction unit and the relationship extraction unit share the characteristic representation generated by the characteristic representation generation unit to extract the terms and the relationships. The term-relationship extraction model is
5. The knowledge model creation support device according to claim 3, wherein, as a learning phase, learning is performed by a multi-task learning process that updates generation parameters of the feature representation in the feature representation generation unit based on a term extraction loss generated by the term extraction unit in extracting the terms and a relationship extraction loss generated by the relationship extraction unit in extracting the relationships. The term-relationship extraction model is
In a first learning phase, learning is performed by a single-task learning process that updates the generation parameters based on the term extraction loss;
6. The knowledge model creation support device according to claim 5 , wherein learning is performed by the multitask learning process as a second learning phase following the first learning phase. The knowledge model creation support device comprises:
a first database, which is the term-relationship database created using the term-relationship extraction model;
a second database that is created using a model different from the term-relationship extraction model and includes relationships between at least a plurality of terms;
Equipped with
The knowledge model creation unit
7. The knowledge model creation support device according to claim 1, wherein the knowledge model is created based on the first database and the second database. The knowledge model creation unit
Draw a knowledge network diagram in a drawing GUI window, the knowledge network diagram being composed of node figures representing the factors and link figures representing relationships between the factors;
A knowledge model creation support device as described in any one of claims 1 to 7, which performs at least one of the following when drawing the knowledge network diagram: displaying placement candidates for the node figures of terms that have a relationship with the factor of interest based on the term-relationship database; displaying terms that have a relationship with the factor of interest based on the term-relationship database as factor candidates; and displaying link figures that have no relationship with the factor of interest based on the term- relationship database.

Images