Knowledge-based dialogue system for self-learning dialogues and learning method thereof
View Patent ↗Disclosed is a self-learning method of a knowledge-based dialogue system. The method includes generating, by the query generating module, a natural language query by itself based on a knowledge path for a query selected from a knowledge graph, inferring, by the answer generating module, a knowledge path for an answer corresponding to a query intention of the natural language query from the knowledge graph, and generating a natural language answer by itself based on the knowledge path for an answer, comparing and evaluating, by the dialogue evaluating unit, the knowledge path for a query and the knowledge path for an answer, and learning, by the query generating module and the answer generating module, a dialogue between the query generating module and the answer generating module based on an evaluation result of the dialogue evaluating unit.
1 . A self-learning method of a knowledge-based dialogue system, as a learning method of a dialogue system including a query generating module, an answer generating module, and a dialogue evaluating unit executed by a processor, the self-learning method comprising:
generating, by the query generating module, a natural language query by itself based on a knowledge path for a query selected from a knowledge graph;
inferring, by the answer generating module, a knowledge path for an answer corresponding to a query intention of the natural language query from the knowledge graph, and generating a natural language answer by itself based on the knowledge path for an answer;
comparing and evaluating, by the dialogue evaluating unit, the knowledge path for a query and the knowledge path for an answer; and
learning, by the query generating module and the answer generating module, a dialogue between the query generating module and the answer generating module based on an evaluation result of the dialogue evaluating unit;
wherein
the generating of the natural language query by itself includes:
selecting a knowledge path for a query connecting an entity used in a current dialogue turn and a correct answer entity to be used in a next dialogue turn in the knowledge graph; and
generating the natural language query corresponding to the knowledge path for a query by using a machine learning model trained to generate the natural language query based on a knowledge graph for learning extracted from a dialogue corpus; and
wherein
the generating of the natural language answer by itself includes:
analyzing the query intention by analyzing a relationship between entities included in the natural language query;
inferring the knowledge path for an answer describing a relationship between at least one entity among the entities in the knowledge graph and a correct answer entity corresponding to the query intention; and
generating a natural language answer based on the correct answer entity and the knowledge path for an answer.
2 . The self-learning method of claim 1 , further comprising:
expressing, by a knowledge expressing unit executed by the processor, complex knowledge including unstructured knowledge and structured knowledge as the knowledge graph including a set of triples describing a relationship between entities.
3 . The self-learning method of claim 1 , wherein
the generating of the natural language query by itself includes:
selecting a correct answer entity to be used in a next dialogue turn based on a partial knowledge graph used in a current dialogue turn, the partial knowledge graph representing a part of the knowledge graph;
selecting the knowledge path for a query describing a relationship between an entity included in the partial knowledge graph and the correct answer entity; and
generating the natural language query based on the knowledge path for a query.
4 . The self-learning method of claim 3 , wherein
the selecting of the correct answer entity includes:
selecting a plurality of candidate correct answer entities connected to the entity included in the partial knowledge graph within the knowledge graph; and
selecting, as the correct answer entity, a candidate correct answer entity having a lowest number of use in the current dialogue turn from among the plurality of candidate correct answer entities.
5 . The self-learning method of claim 1 , wherein
the knowledge path for a query describes a first relationship between a first entity used in the current dialogue turn and a first correct answer entity connected to the first entity in the knowledge graph, and
the knowledge path for an answer describes a second relationship between a second entity included in the natural language query and a second correct answer entity corresponding to the query intention of the natural language.
6 . The self-learning method of claim 5 , wherein
the evaluating includes:
comparing the first correct answer entity with the second correct answer entity; and
comparing the first relationship with the second relationship.
7 . A knowledge-based dialogue system for self-learning conversations, the knowledge-based dialogue system comprising:
a communication interface configured to receive complex knowledge including unstructured knowledge and structured knowledge;
a knowledge conversion unit configured to convert the complex knowledge into a knowledge graph including a set of knowledge triples describing relationships between entities;
a query generating module configured to generate a natural language query by itself based on a knowledge path for a query selected from the knowledge graph;
an answer generating module configured to infer a knowledge path for an answer corresponding to a query intention of the natural language query from the knowledge graph, and generating a natural language answer by itself based on the knowledge path for an answer;
a dialogue evaluating unit configured to compare the knowledge path for a query and the knowledge path for an answer to evaluate a difference between the knowledge path for a query and the knowledge path for an answer; and
a processor configured to control operations of the knowledge conversion unit, the query generating module, the answer generating module, and the dialogue evaluating unit;
wherein generating the natural language query includes using a machine learning model trained to generate the natural language query based on the knowledge graph for learning extracted from a dialogue corpus;
wherein
the generating of the natural language answer by itself includes:
analyzing the query intention by analyzing a relationship between entities included in the natural language query; and
wherein
inferring the knowledge path for an answer includes describing a relationship between at least one entity among the entities in the knowledge graph and a correct answer entity corresponding to the query intention; and
generating a natural language answer based on the correct answer entity and the knowledge path for an answer.
8 . The knowledge-based dialogue system of claim 7 , wherein
the query generating module includes:
a correct answer selecting unit configured to select a correct answer entity to be used in a next dialogue turn based on a partial knowledge graph used in a current dialogue turn, the partial knowledge graph representing a part of the knowledge graph;
a path selecting unit configured to select the knowledge path for a query describing a relationship between an entity included in the partial knowledge graph and the correct answer entity; and
a natural language query generating unit configured to generate the natural language query based on the knowledge path for a query.
9 . The knowledge-based dialogue system of claim 8 , wherein
the correct answer selecting unit selects a plurality of candidate correct answer entities connected to the entity included in the partial knowledge graph in the knowledge graph and then selects a candidate correct answer entity having a least number of use in the current dialogue turn, among the plurality of candidate correct answer entities, as the correct answer entity.
10 . The knowledge-based dialogue system of claim 7 , wherein
the answer generating module includes:
a query understanding unit configured to analyze a relationship between the entities included in the natural language query and analyze the query intention;
a correct answer interfering unit configured to infer the knowledge path for an answer describing a relationship between at least one of the entities in the knowledge graph and the correct answer entity corresponding to the query intention; and
a natural language answer generating unit configured to generate a natural language answer based on the correct answer entity and the knowledge path for an answer.
11 . The knowledge-based dialogue system of claim 7 , wherein
the dialogue evaluating unit compares the knowledge path for a query describing a first relationship between the entity used in the current dialogue turn and a first correct answer entity selected based on the entity used in the current dialogue turn with the knowledge path for an answer describing a second relationship between the entity included in the natural language query and a second correct answer entity inferred based on the entity included in the natural language query.
12 . The knowledge-based dialogue system of claim 11 , wherein
the dialogue evaluating unit compares the first correct answer entity with the second correct answer entity and compares the first relationship with the second relationship.
13 . A self-learning method of a knowledge-based dialogue system, as a learning method of a dialogue system including a query generating module, an answer generating module, and a dialogue evaluating unit executed by a processor, the self-learning method comprising:
selecting, by the query generating module, a knowledge path for a query describing a first relationship between a first entity used in a current dialogue turn and a first correct answer entity to be used in a next dialogue turn in a knowledge graph and generating a natural language query based on the knowledge path for a query;
selecting, by the answer generating module, a knowledge path for an answer describing a second relationship between a second entity included in the natural language query and a second correct answer entity corresponding to a query intention of the natural language query in the knowledge graph and generating a natural language answer by itself based on the knowledge path for an answer;
evaluating, by the dialogue evaluating unit, an error between the knowledge path for a query and the knowledge path for an answer; and
learning, by the query generating module and the answer generating module, the knowledge path for a query and the knowledge path for an answer based on the error,
wherein generating the natural language query includes using a machine learning model trained to generate the natural language query based on the knowledge graph for learning extracted from a dialogue corpus; and
wherein
the generating of the natural language answer by itself includes:
analyzing the query intention by analyzing a relationship between entities included in the natural language query,
inferring the knowledge path for an answer describing a relationship between at least one entity among the entities in the knowledge graph and a correct answer entity corresponding to the query intention; and
generating a natural language answer based on the correct answer entity and the knowledge path for an answer.
14 . The self-learning method of claim 13 , wherein
the knowledge graph expresses a complex knowledge including unstructured knowledge and structured knowledge as a set of triples describing a relationship between entities.
15 . The self-learning method of claim 13 , wherein
the knowledge path for a query is triple data structured in the order of the first entity, the first relationship, and the first correct answer entity, and
the knowledge path for an answer is triple data structured in the order of the second entity, the second relationship, and the second correct answer entity.
16 . The self-learning method of claim 15 , wherein
measuring of an error includes:
measuring an error between the first relationship and the second relationship; and
measuring an error between the first correct answer entity and the second correct answer entity.