Adversarial learning framework for persona-based dialogue modeling
Various embodiments may be generally directed to the use of an adversarial learning framework for persona-based dialogue modeling. In some embodiments, automated multi-turn dialogue response generation may be performed using a persona-based hierarchical recurrent encoder-decoder-based generative adversarial network (phredGAN). Such a phredGAN may feature a persona-based hierarchical recurrent encoder-decoder (PHRED) generator and a conditional discriminator. In some embodiments, the conditional discriminator may include an adversarial discriminator that is provided with attribute representations as inputs. In some other embodiments, the conditional discriminator may include an attribute discriminator, and attribute representations may be handled as targets of the attribute discriminator. The embodiments are not limited in this context.
1. A computer-implemented method, comprising:
receiving, by at least one server communicatively coupled with a user device, a dialogue utterance;
applying, by the at least one server, a generative adversarial network (GAN) to the dialogue utterance to generate response candidates to the utterance and determine a response from the response candidates to respond to the utterance, the GAN comprising a generator and a discriminator, and wherein the applying the GAN comprises:
generating, by the generator, utilizing source attributes and target attributes, the response candidates responsive to the dialogue utterance, wherein the source attributes comprise a speaker identity, a speaker background, a speaker location, a speaker preference, a speaker sentiment, or combination thereof, and the target attributes comprise a respondent identity, a respondent background, a respondent location, a respondent preference, a respondent sentiment, or a combination thereof;
determining, by the discriminator, the response to respond to the dialogue utterance from the response candidates based on discrimination metrics comprising human-likeness and persona, the discriminator comprising an attribute discriminator to utilize the target attributes as a discriminator target and a dialogue history for multi-label attribute classification to classify outputs from the generator with an attribute class, and an adversarial discriminator to determine a binary output for human-likeness to identify the response candidates as real or fake; and
causing, by the at least one server, communication of the response to the user device responsive to the dialogue utterance.
2. The computer-implemented method of claim 1 , wherein the discriminator comprises an utterance-level discriminator.
3. The computer-implemented method of claim 1 , wherein the discriminator comprises a word-level adversarial discriminator.
4. The computer-implemented method of claim 1 , the generator further comprising a context recurrent neural network (cRNN), an encoder RNN (eRNN), an attention RNN (aRNN), and a decoder RNN (dRNN), and the cRNN to concatenate the source attributes with an output of the eRNN to generate an initial state for the dRNN.
5. The computer-implemented method of claim 4 , the dRNN to concatenate the target attributes with an output of the aRNN and generate a decoder output based on the initial state.
6. The computer-implemented method of claim 4 , wherein output of the aRNN is coupled to the dRNN using an additive attention mechanism.
7. The computer-implemented method of claim 4 , wherein the eRNN is a bidirectional RNN and the cRNN is a unidirectional RNN, the eRNN and the cRNN each comprising units, and each unit comprising a three-layer gate recurrent unit (GRU) with a hidden state size of 512.
8. A system, comprising:
a networking interface;
memory to store instructions; and
processing circuitry, coupled with the memory and the networking interface, operable to execute the instructions, that when executed, cause the processing circuitry to:
detect a dialogue utterance received from a user device via the networking interface;
process the dialogue utterance utilizing a generative adversarial network (GAN) logic to generate response candidates to the dialogue utterance, wherein the GAN comprises generator logic and discriminator logic, and processing the dialogue utterance utilizing the GAN logic to:
generate, by the generator logic, utilizing source attributes and target attributes, the response candidates responsive to the dialogue utterance, wherein the source attributes comprise attributes of a speaker, and the target attributes comprise attributes of a respondent; and
determine, by the discriminator logic, the response to respond to the dialogue utterance from the response candidates based on discrimination metrics comprising human-likeness and persona, the discriminator logic comprising an attribute discriminator to utilize the target attributes as a discriminator target and a dialogue history for multi-label attribute classification to classify outputs from the generator with an attribute class, and an adversarial discriminator to determine a binary output for human-likeness to identify the response candidates as real or fake; and
cause, via the networking interface, communication of the response to the user device responsive to the dialogue utterance.
9. The system of claim 8 , wherein the discriminator logic comprises an utterance-level discriminator.
10. The system of claim 8 , wherein the discriminator logic comprises a word-level adversarial discriminator.
11. The system of claim 8 , the generator logic further comprising a context recurrent neural network (cRNN), an encoder RNN (eRNN), an attention RNN (aRNN), and a decoder RNN (dRNN), and the cRNN to concatenate the source attributes with an output of the eRNN to generate an initial state for the dRNN.
12. The system of claim 11 , wherein output of the aRNN is coupled to the dRNN using an additive attention mechanism.
13. The system of claim 11 , wherein the eRNN is a bidirectional RNN and the cRNN is a unidirectional RNN, the eRNN and the cRNN each comprising units, and each unit comprising a three-layer gate recurrent unit (GRU) with a hidden state size of 512.
14. The system of claim 8 , the discriminator logic to determine the response, wherein an adversarial loss is cGAN adv (G, D adv )= X i ,C i+1 ,X i+1 [log D adv (X i , C i+1 , X i+1 )]+ X i , Z i [1−log D adv (X i , G(X i , C i+1 , Z i ))], and wherein an optimization objective for the response is
G
*
,
D
adv
*
,
D
att
*
=
arg
min
G
(
(
max
D
adv
(
λ
G
adv
ℒ
c
GAN
adv
(
G
,
D
adv
)
+
min
D
att
λ
G
att
ℒ
cGAN
att
(
G
,
D
att
)
+
λ
M
ℒ
MLE
(
G
)
)
and λ G att is set to 0, wherein G represents the generator logic, D adv represents an adversarial discriminator, X i represents a sequence of utterances, C i+1 represents the target attributes, X i+1 represents a ground truth, λ G adv represents an adversarial hyperparameter, λ G att represents an attribute prediction hyperparameter, cGAN att represents an attribute prediction loss, D att represents an attribute discriminator, λ M represents a word token hyperparameter, MLE (G) represents a maximum likelihood estimation (MLE) loss, and Z i represents a noise sample.
15. A non-transitory computer-readable medium storing instructions that, when executed, cause one or more processors of one or more servers to:
receive, from a user device, a dialogue utterance;
generate, by generator logic of a generative adversarial network (GAN), utilizing source attributes and target attributes, response candidates responsive to the dialogue utterance, wherein the source attributes comprise attributes of a speaker, and the target attributes comprise attributes of a respondent;
determine, by discriminator logic comprising a discriminator, a response to respond to the dialogue utterance from the response candidates based on discrimination metrics comprising human-likeness and persona, the discriminator comprising an attribute discriminator to utilize the target attributes as a discriminator target and a dialogue history for multi-label attribute classification to classify outputs from the generator with an attribute class, and an adversarial discriminator to determine a binary output for human-likeness to identify the response candidates as real or fake; and
cause, via a networking interface, communication of the response to the user device responsive to the dialogue utterance.
16. The non-transitory medium of claim 15 , wherein the discriminator comprises an utterance-level discriminator.
17. The non-transitory medium of claim 15 , wherein the discriminator comprises a word-level adversarial discriminator.
18. The non-transitory medium of claim 15 , the generator logic further comprising a context recurrent neural network (cRNN), an encoder RNN (eRNN), an attention RNN (aRNN), and a decoder RNN (dRNN), and the cRNN to concatenate the source attributes with an output of the eRNN to generate an initial state for the dRNN.
19. The non-transitory medium of claim 18 , wherein the eRNN is a bidirectional RNN and the cRNN is a unidirectional RNN, the eRNN and the cRNN each comprising units, and each unit comprising a three-layer gate recurrent unit (GRU) with a hidden state size of 512.
20. The non-transitory medium of claim 15 , the discriminator to determine the response, wherein an adversarial loss is cGAN adv (G, D adv )= X i ,C i+1 ,X i+1 [log D adv (X i ,C i+1 ,X i+1 )]+ X i ,Z i [1−log D adv (X i ,G(X i ,C i+1 ,Z i ))], and wherein an optimization objective for the response is
G
*
,
D
adv
*
,
D
att
*
=
arg
min
G
(
(
max
D
adv
λ
G
adv
ℒ
c
GAN
adv
(
G
,
D
adv
)
+
min
D
att
λ
G
att
ℒ
cGAN
att
(
G
,
D
att
)
+
λ
M
ℒ
MLE
(
G
)
)
where λ G att is set to 0, wherein G represents the generator logic, D adv represents an adversarial discriminator, X i represents a sequence of utterances, C i+1 represents the target attributes, X i+1 represents a ground truth, λ G adv represents an adversarial hyperparameter, λ G att represents an attribute prediction hyperparameter, cGAN att represents an attribute prediction loss, D att represents an attribute discriminator, λ M represents a word token hyperparameter, MLE (G) represents a maximum likelihood estimation (MLE) loss, and Z i represents a noise sample.