JP2020191009A - Knowledge graph complementing device and knowledge graph complementing method - Google Patents

Abstract

To provide a knowledge graph complementing device and a knowledge graph complementing method that can improve the accuracy of knowledge graph complementation.SOLUTION: A knowledge graph complementing device 1 comprises: a knowledge graph storage unit 13 that stores a knowledge graph with a graph structure; an acquisition unit (sound collection unit 11) that acquires an utterance; and a complementation unit 15 that, when there is an unknown entity in the acquired utterance, develops the utterance into an embedded representation considering an identification character string based on the pattern of an existing graph, generates a triple candidate for the unknown entity based on a result of the development, calculates the certainty factor of the triple candidate, and acquires knowledge on the unknown entity. The complementation unit 15 generates partial character strings in the identification character strings of the existing entity and the unknown entity as pseudo entities, generates a pseudo triple in which the relation between the existing entity or the unknown entity and the pseudo entities is formed, and calculates the certainty factor of the triple candidate based on the set of a triple and the pseudo triple.SELECTED DRAWING: Figure 1

Description

The present invention relates to a knowledge graph complementing device and a knowledge graph complementing method.

In the dialogue system, a question is asked to a user and information is obtained (see, for example, Patent Document 1). If the required information is not on the knowledge base, current dialogue systems may change topics or respond unnaturally in context, such as "I don't know." In this way, a type of dialogue system that refers to a database and responds cannot respond well to information that is not explicitly described in the database. Therefore, it is desired to construct a dialogue system that can acquire information that is not in the database.

On the other hand, the inventors of the present invention have decided to construct a dialogue system capable of responding, which was difficult in the past, by inferring potential information from a knowledge graph which is a knowledge base having a graph structure. It has been proposed (see Non-Patent Document 1).

Japanese Patent Application No. 2018-062055

Yuma Fujioka, Katsuhiko Hayashi, Mikio Nakano, Kazunori Komatani, "Trial of Response Generation Using Embedded Representation of Knowledge Graph in Dialogue System", Japanese Society for Artificial Intelligence Research Materials, Japanese Society for Artificial Intelligence, 2018, SIG-SLUD, Vol. B5 , No. 02, pp. 88-89

However, in such a conventional system, it is desired to further improve the accuracy of knowledge graph complementation.

The present invention has been made in view of the above problems, and an object of the present invention is to provide a knowledge graph complementing device and a knowledge graph complementing method capable of improving the accuracy of knowledge graph complementation.

(1) In order to achieve the above object, the knowledge graph complement device <1> according to one aspect of the present invention includes a triple with an identification character string of each of the two entities and a label indicating the relationship between the entities. A knowledge graph storage unit <13> that stores a knowledge graph, which is a set and is a database having a graph structure, and an acquisition unit <sound collection unit 11, voice recognition unit 12, language understanding unit 14> that acquires speech. If there is an unknown entity that is not stored in the knowledge graph storage unit in the acquired speech, an embedded expression that considers the identification character string based on the known graph pattern stored in the knowledge graph storage unit. Expand to, generate triple candidates for the unknown entity based on the expanded result, calculate the certainty of the triple candidate, and acquire knowledge about the unknown entity based on the calculated certainty. A part <15> is provided, and the complementary part generates a substring of an existing entity stored in the knowledge graph storage unit and an identification character string of the unknown entity as a pseudo entity, and the existing entity. Alternatively, a pseudo triple formed by adding a label indicating that the relationship between the unknown entity and the pseudo entity is a substring is generated, and based on the triple and the set of pseudo triples, the triple candidate Calculate the certainty.

(2) Further, the complementary unit generates a question about a triple candidate whose conviction rank is higher than a predetermined rank and higher than a predetermined value, or the conviction rank is higher than a predetermined rank. Or generate a question about triple candidates higher than a given value
If the answer to the question is affirmative, the information about the unknown entity may be finalized.

(3) Further, in the knowledge graph complement device according to one aspect of the present invention, the complement unit is
The conviction may be calculated by ComplEx, which is a model using the complex number representation and the Hermitian inner product.

(4) In order to achieve the above object, the knowledge graph complementing method according to one aspect of the present invention is a knowledge graph complementing method in a knowledge graph which is a database having a graph structure, and the knowledge graph is included in the knowledge graph. A set whose elements are triples with identification strings of each of the two entities and a label indicating the relationship between these entities, and the knowledge graph complement device generates a pseudo-entity based on a substring of the identification character string of the entity. At the same time, a pseudo triple is formed with a label indicating that the relationship between the entity and the pseudo entity is a substring, and the pseudo triple is added to the knowledge graph which is a set of triples between the two entities. A second knowledge graph is formed, and for the second knowledge graph, the certainty of an unknown triple regarding an entity is calculated according to the knowledge graph embedding model, and the certainty rank is high and the certainty is equal to or higher than a predetermined value. Is estimated and supplemented as new information, or those with a high degree of certainty or a certainty level higher than a predetermined value are estimated and supplemented as new information.

(5) Further, in the knowledge graph complementing method according to one aspect of the present invention, when a question about the new information is asked to a person and the answer to the question is affirmative.
The new information may be complemented as finalized information.

According to (1) or (4) described above, the accuracy of knowledge graph complementation can be improved.

Further, according to (2) described above, knowledge acquisition can be attempted by asking the user about triple candidates that are not on the knowledge graph. Also, prevent the user from asking uncertain questions even when the certainty is low, that is, when the estimation result is uncertain, and do not discourage the user from interacting with questions that are clearly wrong. can do.
Further, according to (3) described above, the existence likelihood of triples can be appropriately calculated.

It is a figure which shows the configuration example of the knowledge graph complementing apparatus which concerns on embodiment. It is a figure which shows the example of the knowledge graph which the knowledge graph storage part of embodiment stores. It is a figure which shows the example which the information of the knowledge graph is not acquired. It is a figure which shows the example of the dialogue which performs the knowledge acquisition which concerns on embodiment, and the outline of the processing performed by the knowledge graph complement device. It is a figure which shows the development example at N = 2 with respect to "Maple bowl" and "parent-child bowl" which concerns on embodiment. This is an example of giving confidence to the question list. It is a flowchart of the knowledge acquisition process performed by the knowledge graph complement device which concerns on embodiment. It is a figure which shows the accuracy index and learning data and the verification result in each setting in the verification.

Hereinafter, embodiments of the present invention will be described with reference to the drawings. In the drawings used in the following description, the scale of each member is appropriately changed in order to make each member recognizable.

[Configuration of Knowledge Graph Complementary Device 1]
FIG. 1 is a diagram showing a configuration example of the knowledge graph complement device 1 according to the present embodiment.
As shown in FIG. 1, the knowledge graph complement device 1 includes a sound collection unit 11 (acquisition unit), a voice recognition unit 12 (acquisition unit), a knowledge graph storage unit 13 (acquisition unit), a language understanding unit 14, and a complement unit 15. , A dialogue generation unit 16, a scenario storage unit 17, and an output unit 18 (dialogue generation unit).

The knowledge graph complement device 1 interacts with the user and complements information that is not in the knowledge graph to be stored. The knowledge graph will be described later.

The sound collecting unit 11 is a microphone, collects a voice signal, and outputs the collected voice signal to the voice recognition unit 12. The sound collecting unit 11 may be a microphone array composed of a plurality of microphones.

The voice recognition unit 12 performs voice recognition processing on the voice signal output by the sound collection unit 11. The voice recognition process is, for example, a sound source localization process, a noise suppression process, a sound source identification process, a sound source separation process, or the like. The voice recognition unit 12 outputs the recognized result to the language understanding unit 14 in a text format, for example.

The knowledge graph storage unit 13 stores the knowledge graph. The knowledge graph storage unit 13 may be placed on the network. Here, the knowledge graph is shown by the relationship of the connection of entities (Entity). Further, the knowledge graph storage unit 13 stores a threshold value used for comparison with the certainty degree described later.

The language understanding unit 14 understands the utterance content by referring to the knowledge graph stored in the knowledge graph storage unit 13 with respect to the text output by the voice recognition unit 12. When the language understanding unit 14 can understand the text output by the voice recognition unit 12, that is, when the information exists in the knowledge graph, the language understanding unit 14 outputs the understood result to the dialogue generation unit 16. When the language understanding unit 14 cannot understand the text output by the speech recognition unit 12, that is, when the information does not exist in the knowledge graph, the language understanding unit 14 outputs the word for which there is no information in the knowledge graph to the complement unit 15. Further, when the user's response to the question sentence indicates affirmative content, the language understanding unit 14 outputs a complementary instruction for complementing the knowledge graph to the complementary unit 15 based on the question content. When the user's response to the question sentence is negative, the language understanding unit 14 outputs information indicating that the user could not understand the question to the dialogue generation unit 16.

The complement unit 15 acquires a word for which information does not exist in the knowledge graph output by the language understanding unit 14. The complement unit 15 generates question information for the acquired word, and outputs the generated question information to the dialogue generation unit 16. The method of generating question information will be described later. The complement unit 15 stores the knowledge graph in the knowledge graph so as to complement the knowledge graph based on the question content based on the complement instruction output by the language understanding unit 14.

The dialogue generation unit 16 generates a question sentence based on the question information output by the complement unit 15 and the language model stored in the scenario storage unit 17, and outputs the generated question sentence to the output unit 18. Alternatively, for example, a dialogue sentence for continuing the dialogue is generated based on the understanding result output by the language understanding unit 14 and the scenario information, and the generated dialogue sentence is output to the output unit 18.

The scenario storage unit 17 stores a scenario according to the dialogue scene. The scenario storage unit 17 also stores the language model.

The output unit 18 is a speaker. The output unit 18 reproduces the audio signal output by the dialogue generation unit 16.

[Example of knowledge graph]
Next, an example of the knowledge graph (see Reference 1) will be described.
FIG. 2 is a diagram showing an example of a knowledge graph stored in the knowledge graph storage unit 13 of the present embodiment. In the example shown in FIG. 2, the word stored in the knowledge graph is an example of "Oyakodon". Reference numerals g101, g141, g151 to g154, and g161 are nodes. Reference numerals g111, g121-124, and g131 are edges. For example, the edge of reference numeral g111 is taste. The edge of reference numeral g121 is a material. The edge of the symbol g131 is a dish type. In this way, the taste, ingredients, and cooking types of "Oyakodon" have been acquired. The knowledge graph shown in FIG. 2 is an example, and the structure of the knowledge graph is not limited to this.

Here, an example in which knowledge has not been acquired in the knowledge graph will be described.
FIG. 3 is a diagram showing an example in which the information of the knowledge graph is not acquired. In the example shown in FIG. 3, the word is "Maple bowl". Reference numerals g101a, g141a, g151a to g153a, and g161a are nodes. Reference numerals g111a, g121a to 123a, and g131a are edges. For example, the edge of reference numeral g111a is taste. The edge of reference numeral g121a is a material. The edge of the symbol g131a is a dish type. In this way, for "Maple bowl", the taste and ingredients have been acquired, but the cooking type has not been acquired.

Reference 1; Angles, R. and Gutierrez, C., “Survey of Graph Database Models”, ACM Comput, Surv., 2008, Vol. 40, No. 1, pp. 1: 1-1: 39

[Framework for dialogue to acquire knowledge]
Next, an example of a dialogue for acquiring knowledge and an outline of the processing performed by the knowledge graph complement device 1 will be described.
FIG. 4 is a diagram showing an example of a dialogue for acquiring knowledge according to the present embodiment and an outline of processing performed by the knowledge graph complement device 1.

First, the knowledge graph complement device 1 acquires the "I ate Momiji-don for the first time, but it was delicious" that the user said.

The knowledge graph complementing device 1 searches the knowledge graph storage unit 13 and recognizes that the information regarding the "maple bowl" is insufficient. In this case, as shown in FIG. 3, it is assumed that the information on the cooking type is insufficient.

Next, the knowledge graph complement device 1 generates a question list based on the relationship between "cooking types" having unknown information.

Next, the knowledge graph complement device 1 converts it into an embedded representation based on a known graph pattern. The embedded expression will be described later.

Next, the knowledge graph complementing device 1 calculates a score representing the existence likelihood for the triple in the question list from the embedded expression and gives it as a certainty.

Next, the knowledge graph complement device 1 selects a question with the highest degree of certainty and the degree of certainty is equal to or higher than the threshold value, and asks the user the question "Is Momiji-don Japanese food?" Try to acquire knowledge. In this embodiment, not only the ranking by the conviction but also the absolute value of the conviction is considered by the threshold value. The reason for this is to prevent the user from asking the user even when the certainty is low, that is, when the estimation result is uncertain, and to prevent the user from being discouraged from interacting with a question that is clearly wrong. Is.

Next, in the knowledge graph complementing device 1, since the user's utterance "That's right!" Expresses affirmation, the knowledge graph acquires the fact that the cooking type of "Maple bowl" is "Japanese food" and stores it in the knowledge graph. To do.

As described above, in the present embodiment, when there is information that is not in the knowledge graph in the audio signal emitted by the user, a question for acquiring the information is generated, and the knowledge graph is complemented based on the question. To do.

[How to complement the knowledge graph]
Next, an example of how to complement the knowledge graph will be described.
As shown in FIG. 4, the complement 15 generates a question list in order to acquire triples from the user.

Knowledge graphs are generally represented as labeled directed graphs. The edges on the directed graph are labeled to represent the relationship. Nodes correspond to entities. Let ε be the set of entities on the knowledge graph and R be the set of relations. For e _i , e _k ∈ ε, ω _j ∈ R, we define a triplet (i, j, k) as a triple. At this time, i, j, and k are defined as the subject, predicate, and object, respectively. A triple is a basic element of a knowledge graph that represents the relationship between two entities. The knowledge graph G can be represented as a set having triples as elements.

In the example shown in FIG. 4, the information on the cooking type is not acquired for "Maple bowl". For this reason, the complementary unit 15 generates a question list so as to fill in the triples (maple bowl, cooking type,?) And the object. The question list generated in this embodiment is generated in a format that fills in the object as in this example. In the question, the complement unit 15 refers to the information stored in the scenario storage unit 17 to generate a question list related to the context of the dialogue and the topic, and avoids a sudden topic change.

[Embedding knowledge graph using identification character string]
Next, the complementary unit 15 converts it into an embedded expression (see Reference 2) based on a known graph pattern. The model learns embedded expressions based on how entities are linked to each other, but the information about the identification character strings (names) of all entities is not incorporated into the learning. Therefore, in the present embodiment, the identification character string of the entity on the knowledge graph is decomposed into subcharacter strings, and the first or last few characters are created and incorporated as a pseudo entity. Under the assumption that entities with similar substrings are likely to have similar properties, the relationships between such entities will be stronger, leading to improved complement accuracy. The embedded expression is an embedding in a latent space.

In the present embodiment, it is described that a pseudo entity is created with a substring at the beginning or the end of N characters or less to form a new triple, which is expanded with N characters. FIG. 5 is a diagram showing an example of development at N = 2 for the “maple bowl” and the “parent-child bowl” according to the present embodiment. In addition, FIG. 5 is an example of expansion with two characters.

The complementary unit 15 extracts one character and two characters at the beginning and end of the “maple bowl” and the “parent-child bowl” that are the targets of expansion, respectively. Complementary unit 15 extracts the first character "mo", the first two characters "fir", the last two characters "jidon", and the last one character "don" from the "maple bowl". To do. Complementary unit 15 extracts the first character "parent", the first two characters "parent and child", the last two characters "child bowl", and the last character "don" for the "parent-child bowl". To do. In the present embodiment, the extracted first 1 and 2 characters and the last 1 and 2 characters are referred to as pseudo entities.

Further, the complementing unit 15 creates an entity (pseudo-entity) having an identifier with a subs indicating that it is a sub-character string at the beginning of the expanded first 1, 2 characters and last 1 or 2 characters. As a result, "subs_mo", "subs_fir", "subs_jidon", "subs_don", "subs_parent", "subs_child", "subs_child bowl", "subs_don" are created as pseudo-entities. Will be done.

Here, in FIG. 5, the label "prefix_1" represents the first character. The label "prefix_2" represents the first two characters. The label "sufix_1" represents the last character. The label "sufix_2" represents the last two characters. Then, a triple is formed by the relation of prex_N or suffix_N toward the pseudo entity created from the entity of the expansion source. In the example of FIG. 5, the two dishes are connected via a pseudo-entity called "subs_don". In the present embodiment, such expansion is referred to as knowledge graph embedding using a substring.

In the example shown in FIG. 5, N = 2, that is, an example of 2 characters has been described, but N may be 3 or more (3 or more characters). For example, the complementary unit 15 has one character from the beginning "ta", two characters from the beginning "tara", three characters from the beginning "tarako", and three characters from the end for "tarako chazuke" whose cooking type is unknown. You may extract the character "chazuke", the last two characters "pickled", and the last one character "ke". Further, in the complementary unit 15, for "Nori Chazuke" whose cooking type is known, one character from the beginning is "no", two characters from the beginning is "Nori", three characters from the beginning are "Nori Cha", and from the end. You may extract 3 characters as "chazuke", 2 characters from the end as "pickled", and 1 character from the end as "ke".

Reference 2; Kadlec, R., Bajgar, O., and Kleindienst, J., “Knowledge Base Completion: Baselines Strike Back”, 2017, in Proceedings of the 2nd Workshop on Representation Learning for NLP, pp. 69-74

[Knowledge Graph Embedding and ComplexEx]
The calculation of the existence likelihood of triples uses the embedded representation of the knowledge graph. Knowledge graph embedding is known as a typical analysis method for the knowledge graph. By embedding a knowledge graph in a low-dimensional linear space and generalizing it, the presence or absence of missing links on the graph is inferred and complemented. The knowledge graph can be expressed as a third-order tensor X of | ε | × | R | × | ε |, and the (i, j, k) elements x _{i, j, k of} X are of the following equation (1). It is expressed as.

Using this expression, the probability P (x _{i, j, k} = 1) that a triple (i, j, k) exists on the knowledge graph in the knowledge graph embedding is calculated by using the score function φ corresponding to the model. It is expressed as in equation (2).

In equation (2), σ (・) is a sigmoid function, and Θ represents a parameter in each model.
Here, ComplEx (Reference 3) is a kind of knowledge graph embedding model. Note that CompactEx is a matrix that represents the relationship between nodes. ComplEx is known as a model that utilizes the complex representation and the Hermitian dot product. Here, C _n (C is a set of all complex numbers) is a complex n-dimensional space, and e _i , e _k ∈ C ^D , and w _j ∈ C ^D are D-dimensional embedded vectors related to the relationships between entities. Let Re (x) be the real part of x. As a result, the score function φ _{Comp in} ComplEx is expressed by the following equation (3).

Here, when the l-th element of the vector v is vl, it is defined as <a, b, c>: = Σ _k a _k b _{k k k} . ^v_ is the complex conjugate vector of v.

The complementary unit 15 uses the score function obtained in this way as the conviction, and imparts this conviction to the question list. FIG. 6 is an example in which the question list is given a certain degree of certainty. In the example shown in FIG. 6, the certainty of the triple (maple bowl, cooking type, Japanese food) is 0.0015, which is the highest.

Reference 3: Trouillon, T., Welbl, J., Riedel, S., Gaussier, E., and Bouchard, G .: “Complex Embeddings for Simple Link Prediction”, 2016, in Proceedings of the 33rd International Conference on International Conference on Machine Learning, pp. 2071-2080

As described above, in the present embodiment, when there is an unknown entity that is not stored in the knowledge graph storage unit 13 in the acquired utterance, the complementary unit 15 stores the known graph pattern stored in the knowledge graph storage unit 13. Based on, it is expanded to the embedded expression considering the identification character string. Then, the complement unit 15 generates triple candidates for unknown entities based on the expanded result, and calculates the certainty of the generated triple candidates. Then, the complementary unit 15 acquires knowledge about an unknown entity based on the calculated conviction.

For example, if there is no link between "Momiji-don" and "Oyako-don" and "Momiji-don" is an unknown entity, via the pseudo-entity "subs_don" (parent-child bowl, sufix_1, subs_don) The predicate (maple bowl, sufix_1, subs_don) and the object have the same relationship. From this, it is possible to give information to the model that "Maple bowl" and "Oyakodon" may be similar. In the present embodiment, using this information, it is inferred by ComplEx that, for example, "Oyakodon" may have the attribute "Cooking type is Japanese", and "Maple bowl" may also have it. Then, in the present embodiment, when the triple (parent-child bowl, cooking type, Japanese food) is included in the learning data (if known), it does not exist in the learning data (maple bowl, cooking type ,?). Of the triples, the triple with the highest degree of certainty (maple bowl, cooking type, Japanese food) obtained by ComplEx was used to acquire knowledge. Thereby, according to the present embodiment, it is possible to acquire knowledge about an unknown entity.

[Example of processing procedure]
Next, an example of the processing procedure will be described. FIG. 7 is a flowchart of the knowledge acquisition process performed by the knowledge graph complement device according to the present embodiment.

(Step S1) The sound collecting unit 11 collects an audio signal. Subsequently, the voice recognition unit 12 acquires the voice signal.

(Step S2) The voice recognition unit 12 performs voice recognition processing on the acquired voice signal.

(Step S3) The language understanding unit 14 determines whether or not there is information whose relation is unknown in the phrase that has undergone voice recognition processing. When the language understanding unit 14 determines that there is no information whose relation is unknown (step S3; NO), the language understanding unit 14 ends the process. When the language understanding unit 14 determines that there is information whose relation is unknown (step S3; YES), the language understanding unit 14 proceeds to the process of step S4.

(Step S4) Complementary unit 15 generates a question list from a relation having unknown information.

(Step S5) Complementary unit 15 converts unknown information into an embedded representation based on a known graph pattern.

(Step S6) The complementary unit 15 expands the converted embedded expression into embedding a knowledge graph using a sub-character string.

(Step S7) The complementary unit 15 calculates the certainty (score) for each question in the question list based on the expanded result, and gives the question a certainty.

(Step S8) The complement unit 15 selects a triple (subject, predicate, object) having the highest conviction and an absolute value of conviction equal to or greater than a threshold value.

(Step S9) The dialogue generation unit 16 generates a question sentence based on the triple selected by the complement unit 15 and the language model stored in the scenario storage unit 17. Subsequently, the output unit 18 outputs a question sentence.

(Step S10) The sound collecting unit 11 collects an audio signal. Subsequently, the voice recognition unit 12 acquires the voice signal.

(Step S11) The voice recognition unit 12 performs voice recognition processing on the acquired voice signal.

(Step S12) The language understanding unit 14 determines whether or not the utterance is positive as a result of the voice recognition process. When the language understanding unit 14 determines that the utterance was affirmative (step S12; YES), the language understanding unit 14 proceeds to the process of step S13. When the language understanding unit 14 determines that the utterance was negative (step S12; NO), the language understanding unit 14 proceeds to the process of step S14.

(Step S13) The complementary unit 15 stores the information in the knowledge graph, assuming that unknown information has been acquired based on the triple used for the question. After the process, the knowledge acquisition process is terminated.

(Step S14) Since the utterance picked up in step S1 contains unknown information, the dialogue generation unit 16 generates a question sentence for the unknown information based on the scenario stored in the scenario storage unit 17. .. Subsequently, the output unit 18 outputs a question sentence. After the process, the knowledge acquisition process is terminated.

The question text output in step S14 is, for example, if the utterance in step S1 is "I ate Momiji-don for the first time, but it was delicious", "What kind of dish is Momiji-don?" Generate the question text of. Then, the knowledge graph complementing device 1 may acquire unknown information based on the answer to this question when the knowledge cannot be acquired by the storage method of the present embodiment.

[Verification example]
Next, the accuracy of complementation by embedding the knowledge graph using the above-mentioned substring is verified, and the result of confirming the effect will be described.

First, the data used for verification will be described.
The knowledge graph used for verification was created based on a tabular database on cooking for the purpose of operation in an interactive system. In detail, the knowledge graph stores dishes, their ingredients, types of dishes, tastes, places to eat, and so on. This database was created manually and has only partial information. The above method was applied to the knowledge graph created from this database to supplement the information.

This knowledge graph has the number of entities | ε | = 7289, the number of relations | R | = 14, and the number of triples is 22321. The average length of the identification string given to the entity is 5.88 (characters) and the standard deviation is 3.18 (characters).

Next, the verification method will be described.
In the accuracy verification, the knowledge graph for verification was randomly divided into five, and four of them were used as training data G'and one was used as test data H'. In the accuracy verification, training data was embedded, and Hits @ K and MRR, which are accuracy indexes generally used in knowledge graph complementation, were calculated from the embedded expression and H'. Hits @ K calculates the score by replacing k'of (i', j', k'; Θ) with all entities, and in the obtained ranking, φ _Comp (i', j', k'; Θ) refers to the ratio of the top K ranks. MRR is also called the average reciprocal, and is expressed as the average of the reciprocals of the reciprocals of the reciprocals of φ _Comp (i', j', k';

In order to verify the effect of embedding the knowledge graph using substrings, in addition to the case where the training data is expanded with the number of expanded characters N = 1, 2, ..., 7, Hits @ K when the expansion is not performed as a comparison target. The MRR was calculated. Details regarding the training data at the time of deployment are shown in the right part of FIG. The left part of FIG. 8 shows the result of 5-fold cross-validation. FIG. 8 is a diagram showing the accuracy index, the learning data, and the verification result in each setting in the verification.

The number of test data is 4464. In the verification, embedding by CompLEx was applied in common in all cases. The embedded dimension was a complex 200 dimension, and learning was performed by logistic regression. The learning rate was adjusted by using Adagrad following the method of Reference 3, and the number of iterations was 1000 and the number of negative example samplings was 5. In the ranking at the time of index calculation, triples including the expanded substring entities were excluded from the ranking. In the verification, triples already included in the test data were also excluded from the ranking.

Next, the verification result will be described.
As shown in the left part of FIG. 8, it was found that all the indexes were increased with respect to the comparison target (Baseline) regardless of the number of expanded characters N.
In addition, Hits @ 3, 5 and 10 increased in proportion to N. On the other hand, Hits @ 1 took the maximum value when N = 3, and tended to decrease as it increased to N = 4, 5, .... Similarly, for MRR, the time when N = 3 was the maximum.

An example complemented with this divided data will be described. For example, there was a simple complementary example in which substrings such as (Tarako Chazuke, is_a, Ochazuke) were directly linked, and it was improved from 2987th place to 1st place. In addition, is_a is a predicate as described above, and represents "cod roe chazuke" + "is". In addition, there were cases such as (hot cake, taste, sweetness) that were considered to be improved from the information of entities connected by substrings, and were improved from 8th place to 1st place.

Next, the accuracy of each relation was confirmed. For a certain divided data, Hits @ K calculated by the above method was classified and calculated for each relation. It was found that there was no relation in which the accuracy of Hits @ 3, 5 and 10 deteriorated, and that the accuracy was higher than that of the comparison target as a whole. In addition, the accuracy of relations representing hierarchical structures, types, etc. such as "is_a" and "cooking type" was greatly improved with respect to the comparison target. In addition to that, the same increase in accuracy was confirmed for relations such as "taste" and "temperature".

[Question example]
An example of a question that can be generated will be described. The learning method is the same as the verification, but the knowledge graph is not divided and all are used for learning. Two settings when the comparison target and N = 3 are expanded will be described. For example, for the fact (Tsukimi dumpling, taste, sweet), the ranking and its score were calculated using the question list as (Tsukimi dumpling, taste,?). In the comparison target setting, the ranking was 5th and the score was -4.00. When N = 3, the ranking was 1st and the score was -1.68. "Tsukimi dumpling" and "sweet" were not linked by the expansion of the substring, but it is thought that the ranking and score were improved based on the information on the taste of other entities ending with "dumpling". ..

In addition, the ranking and score were calculated in the same way for the fact (rice cooked with five eyes, temperature, warmth). In the comparison target setting, the ranking was 7th and the score was -4.13. When N = 3, the ranking was 1st and the score was -1.66. In this example as well, it can be inferred that the reason for the improvement is that the relevance to the dishes with "rice" at the end has increased.

As described above, in the present embodiment, when an unknown word appears, the accuracy of acquiring the knowledge (attribute) about the word based on the existing knowledge graph (database) is improved. In this embodiment, the first or last few characters of the word (entity or node) character string in the knowledge graph are incorporated into the knowledge graph as pseudo-entities. In the present embodiment, on the assumption that entities having the same substring are likely to have similar properties, the relevance of the entities connected via the pseudo-entity is strengthened and the accuracy is improved.

In the above-mentioned example, an example of acquiring a user's utterance as a voice signal has been described, but the present invention is not limited to this. The utterance to be acquired may be a text file. In this case, the language understanding unit 14 may acquire the text input by the user by operating, for example, a keyboard (not shown) as shown by the chain line in FIG. As a result, the knowledge graph complement device 1 can acquire unknown information even if it is difficult for the user to speak.

The knowledge graph complement device 1 described above can also be applied to, for example, a humanoid robot, a reception system, a response system such as a vehicle, or a response system such as a smartphone. By applying it to these devices and systems, it is possible to efficiently complement the knowledge graph used by these devices and systems and acquire unknown information.

A program for realizing all or part of the functions of the knowledge graph complement device 1 in the present invention is recorded on a computer-readable recording medium, and the program recorded on the recording medium is read into the computer system. By executing it, all the processes performed by the knowledge graph complement device 1 may be performed. The term "computer system" as used herein includes hardware such as an OS and peripheral devices. Further, the "computer system" shall also include a WWW system provided with a homepage providing environment (or display environment). Further, the "computer-readable recording medium" refers to a portable medium such as a flexible disk, a magneto-optical disk, a ROM, or a CD-ROM, or a storage device such as a hard disk built in a computer system. Furthermore, a "computer-readable recording medium" is a volatile memory (RAM) inside a computer system that serves as a server or client when a program is transmitted via a network such as the Internet or a communication line such as a telephone line. In addition, it shall include those that hold the program for a certain period of time.

Further, the program may be transmitted from a computer system in which this program is stored in a storage device or the like to another computer system via a transmission medium or by a transmission wave in the transmission medium. Here, the "transmission medium" for transmitting a program refers to a medium having a function of transmitting information, such as a network (communication network) such as the Internet or a communication line (communication line) such as a telephone line. Further, the above program may be for realizing a part of the above-mentioned functions. Further, a so-called difference file (difference program) may be used, which can realize the above-mentioned functions in combination with a program already recorded in the computer system.

Although the embodiments for carrying out the present invention have been described above using the embodiments, the present invention is not limited to these embodiments, and various modifications and substitutions are made without departing from the gist of the present invention. Can be added.

1 ... Knowledge graph complement device, 11 ... Sound collection unit, 12 ... Speech recognition unit, 13 ... Knowledge graph storage unit, 14 ... Language comprehension unit, 15 ... Complementary unit, 16 ... Dialogue generation unit, 17 ... Scenario storage unit, 18 … Output section

Claims (5)

A set whose elements are triples with identification strings of each of two entities and labels indicating the relationship between the entities.
A knowledge graph storage unit that stores the knowledge graph, which is a database with a graph structure,
The acquisition department that acquires utterances,
If there is an unknown entity that is not stored in the knowledge graph storage unit in the acquired speech, an embedded expression that considers the identification character string based on the known graph pattern stored in the knowledge graph storage unit. Expand to, generate triple candidates for the unknown entity based on the expanded result, calculate the certainty of the triple candidate, and acquire knowledge about the unknown entity based on the calculated conviction. With a department,
The complementary part
A substring of the identification character string of the existing entity and the unknown entity stored in the knowledge graph storage unit is generated as a pseudo entity, and the relationship between the existing entity or the unknown entity and the pseudo entity is established. A pseudo triple formed by adding a label indicating that it is a substring is generated, and the certainty of the triple candidate is calculated based on the triple and the set of pseudo triples.
Knowledge graph complementer. The complementary part
Generate a question about a triple candidate whose confidence rank is higher than a predetermined rank and higher than a predetermined value, or for a triple candidate whose confidence rank is higher than a predetermined rank or higher than a predetermined value. Generate a question and
If the answer to the question is affirmative, the information about the unknown entity is finalized.
The knowledge graph complement device according to claim 1. The complementary part
The conviction is calculated by ComplEx, which is a model using the complex number representation and the Hermitian inner product.
The knowledge graph complement device according to claim 1 or 2. It is a knowledge graph complementation method in the knowledge graph, which is a database with a graph structure.
The knowledge graph is a set whose elements are triples with identification strings of each of the two entities in the knowledge graph and labels indicating the relationship between these entities.
Knowledge graph complementer
A pseudo-entity is generated by a substring of the identification string of the entity, and a pseudo-triple is formed with a label indicating that the entity is a substring as a relationship between the entity and the pseudo-entity, and the two entities A second knowledge graph is formed by adding the pseudo-triple to the knowledge graph, which is a set of triples in between.
For the second knowledge graph, the conviction of an unknown triple regarding an entity is calculated according to the knowledge graph embedding model, and the one with a high conviction rank and a conviction of a predetermined value or more is estimated and complemented as new information. Alternatively, new information is estimated and supplemented with a high degree of certainty or a certainty of more than a predetermined value.
Knowledge graph complement method. If you ask a person a question about the new information and the answer to the question is affirmative,
Complementing the new information as finalized information,
The knowledge graph complement method according to claim 4.

Images