Dialogue model training method and device therefor
Disclosed is a method of training a dialogue model in an electronic device, the method including selecting a first context from a first dialogue data set including at least one pair of a context and a response corresponding to the context, generating a first response corresponding to the first context through a first dialogue model, generating an augmented dialogue dataset by incorporating a pair of the first context and the first response corresponding to the first context into the first dialogue data set, and training a second dialogue model based on the augmented dialogue dataset.
1. A method of training a dialogue model in an electronic device, the method comprising:
selecting a first context from a first dialogue data set including at least one pair of a context and a response corresponding to the context;
generating a first response corresponding to the first context through a first dialogue model, wherein the first dialogue model is a generative-based dialogue model that generates a response to a given context;
generating an augmented dialogue dataset by incorporating a pair of the first context and the first response corresponding to the first context into the first dialogue data set;
generating an augmented response set including a response of the first dialogue data set and the first response; and
training a second dialogue model based on the augmented dialogue dataset, wherein the second dialogue model is a retrieval-based dialogue model that searches for a response to the given context, wherein the training comprises:
acquiring a response set including a first response subset corresponding to a second context included in the augmented dialogue dataset and a second response subset selected arbitrarily,
calculating a first score for a response included in the response set with respect to the second context based on the first dialogue model,
calculating a second score for a response included in the response set with respect to the second context based on the second dialogue model, and
training the second dialogue model based on the first score and the second score.
2. The method of claim 1 , wherein the training of the second dialogue model based on the first score and the second score comprises:
calculating a loss based on the first score and the second score; and
training the second dialogue model such that the loss is minimized.
3. A method of training a dialogue model in an electronic device, the method comprising:
acquiring, from a first dialogue data set, a response set including a first response subset corresponding to a first context and a second response subset selected arbitrarily;
calculating a first score for a response included in the response set with respect to the first context based on a first dialogue model, wherein the first dialogue model is a generative-based dialogue model that generates a response to a given context;
calculating a second score for a response included in the response set with respect to the first context based on a second dialogue model, wherein the second dialogue model is a retrieval-based dialogue model that searches for a response to the given context; and
training the second dialogue model based on the first score and the second score, wherein the training of the second dialogue model includes
calculating a loss based on the first score and the second score; and
training the second dialogue model such that the loss is minimized.
4. The method of claim 3 , further comprising:
selecting a second context from a second dialogue data set including at least one pair of a context and a response corresponding to the context;
generating a response corresponding to the second context through the first dialogue model; and
generating the first dialogue data set by incorporating a pair of the second context and a response corresponding to the second context into the second dialogue data set.
5. The method of claim 3 , wherein the calculating of the second score comprises:
encoding the first context and a response included in the response set based on fixed-length embedding; and
calculating a relevance score with respect to the first context for each response included in the response set based on an embedding value corresponding to the first context and an embedding value corresponding to each response included in the response set.
6. The method of claim 3 , wherein the first score is calculated using a normalized log-likelihood based on a length of each response included in the response set.
7. The method of claim 3 , wherein the first score is calculated based on a mutual information score with respect to the first context for each response included in the response set.
8. The method of claim 3 , wherein the loss comprises a cross-entropy loss for a score corresponding to the first response subset and a knowledge distillation loss for a score corresponding to a response included in the response set.
9. The method of claim 8 , wherein the training of the second dialogue model such that the loss is minimized comprises training by maximizing a score corresponding to the first response subset so that the cross-entropy loss is minimized.
10. The method of claim 8 , wherein the training of the second dialogue model such that the loss is minimized comprises training by matching the first score and the second score so that the knowledge distillation loss is minimized.
11. A non-transitory computer-readable recording medium comprising a computer program to execute the method of claim 3 .
12. An electronic device for training a dialogue model, the electronic device comprising:
a storage device; and
a controller,
wherein the controller is configured to
acquire a response set including a first response subset corresponding to a first context and a second response subset selected arbitrarily, from a first dialogue data set through the storage device;
calculate a first score for a response included in the response set with respect to the first context based on a first dialogue model, wherein the first dialogue model is a generative-based dialogue model that generates a response to a given context;
calculate a second score for a response included in the response set with respect to the first context based on a second dialogue model, wherein the second dialogue model is a retrieval-based dialogue model that searches for a response to the given context; and
train the second dialogue model based on the first score and the second score, wherein, to train the second dialogue model, the controller is configured to
calculate a loss based on the first score and the second score, and
train the second dialogue model such that the loss is minimized.
13. The electronic device of claim 12 , wherein the controller is configured to:
select a second context from a second dialogue data set including at least one pair of a context and a response corresponding to the context;
generate a response corresponding to the second context through the first dialogue model; and
generate the first dialogue data set by incorporating a pair of the second context and a response corresponding to the second context into the second dialogue data set.
14. The electronic device of claim 13 , wherein the controller is configured to:
generate an augmented response set including a response of the second dialogue data set and a response corresponding to the second context; and
store the augmented response set through the storage device.
15. The electronic device of claim 12 , wherein the loss comprises a cross-entropy loss for a score corresponding to the first response subset and a knowledge distillation loss for a score corresponding to a response included in the response set.