Adversarial Learning for Neural Dialogue Generation

Presenter: Keon Kim

Original Paper by: Jiwei Li, Will Monroe, Tianlin Shi, Alan Ritter and Dan Jurafsky

TODOs● What is this about?● Result First!● Why and How on Text Data?● Adversarial Learning?● The Model Breakdown

- Generative- Discriminative

● Training Methods- Monte Carlo Policy Gradient (REINFORCE)- Reward for Every Generation Step (REGS)

● Teacher Forcing● Notes

What Is This About?

- Adversarial Training for open-domain dialogue generation

“to train to produce sequences that are indistinguishable from human-generated dialogue utterances.”

Result First!

Adversarially-Trained system generates higher-quality responses than previous baselines!

Adversarial Training

MinMax Game between Generator vs Discriminator

Why and How on Text Data?

- Analogous to Turing Test ( just discriminator instead of human )

- Enjoyed great success in computer vision

- But hard to apply to NLP because the text space is too discontinuous

- Small updates generally don’t change the reinforcement feedback

- Progress has been made and this is one of them

Given a dialogue history X consisting of a sequence of dialogue utterances, the model needs to generate a response Y. We view the process of sentence generation as a sequence of actions that are taken according to a policy defined by an encoder-decoder recurrent neural networks.

Model Breakdown

The model has two main parts, G and D:

Generative Model (G)- Generates a response y given dialogue history x.- Standard Seq2Seq model with Attention Mechanism

Discriminative Model (D)- Binary Classifier that takes as input a sequence of dialogue

utterances {x, y} and outputs label indicating whether the input is generated by human or machines

- Hierarchical Encoder + 2 class softmax function -> returns probability of the input dialogue episode being a machine or human generated dialogues.

Training Methods (Important Part)

Policy Gradient Methods:

- The score of current utterances being human-generated ones assigned by the discriminator is used as a reward for the generator, which is trained to maximize the expected reward of generated utterances using REINFORCE algorithm.

Uses Monte Carlo Policy Gradient (REINFORCE)

approximated by likelihood ratio

Training Methods (Important Part)

Policy Gradient Methods:

- The score of current utterances being human-generated ones assigned by the discriminator is used as a reward for the generator, which is trained to maximize the expected reward of generated utterances using REINFORCE algorithm.

Uses Monte Carlo Policy Gradient (REINFORCE)

approximated by likelihood ratio

classification scorebaseline value to reduce the variance of the estimate while keeping it unbiased

policygradient in parameter space

Training Methods (Important Part)

Policy Gradient Methods:

- The score of current utterances being human-generated ones assigned by the discriminator is used as a reward for the generator, which is trained to maximize the expected reward of generated utterances using REINFORCE algorithm.

Uses Monte Carlo Policy Gradient (REINFORCE)

approximated by likelihood ratio

scalar reward

policy updates by the direction of the reward in the parameter space

Training Methods (Cont’d)

Problem with REINFORCE:

- has disadvantage that the expectation of the reward is approximated by only one sample

- reward associated with the sample is used for all actions

- REINFORCE assigns the same negative reward to all tokens [I, don’t, know] by comparing them with I don’t know

- Proper credit assignment in training would give separate rewards, - most likely a neutral token for token I, and negative reward to don’t and know.

Authors of the paper calls it: Reward for Every Generation Step (REGS)

Input : What’s your name human : I am Johnmachine : I don’t know

Reward for Every Generation Step (REGS)

We need rewards for intermediate steps.

Two Strategies Introduced:

1. Monte Carlo (MC) Search

2. Training Discriminator For Rewarding Partially Decoded Sequences

Monte Carlo Search

1. Given a partially decoded step s, the model keeps sampling tokens from the distribution until the decoding

finishes

2. Repeats N times (N generated sequences will share a common prefix s).

3. These N sequences are fed to the discriminator, the average score of which is used as a reward.

Rewarding Partially Decoded Sequences

Directly train a discriminator that is able to assign rewards to both fully and partially decoded sequences

- Break generated sequences into partial sequences

Problem:

- Earlier actions in a sequence are shared among multiple training examples for discriminator.

- Result in overfitting

The author proposes a similar strategy used in AlphaGo to mitigate the problem.

Rewarding Partially Decoded Sequences

For each collection of subsequences of Y, randomly sample only one example from positive examples

and one example from negative examples, which are used to update discriminator.

- Time effective but less accurate than MC model.

Rewarding Partially Decoded Sequences

For each collection of subsequences of Y, randomly sample only one example from positive examples

and one example from negative examples, which are used to update discriminator.

- Time effective but less accurate than MC model.

classification score baseline value to reduce the variance of the estimate while keeping it unbiased

policy

gradient in parameter space

Rewarding Partially Decoded Sequences

classification score baseline value

policy

gradient in parameter space

classification scorebaseline value

policygradient in parameter space

Teacher Forcing

Generative model is still unstable, because:

- generative model can only be indirectly exposed to the gold-standard target sequences through the

reward passed back from the discriminator.

- This reward is used to promote or discourage the generator’s own generated sequences.

This is fragile, because:

- Once a generator accidentally deteriorates in some training batches

- And Discriminator consequently does an extremely good job in recognizing sequences from the

generator, the generator immediately gets lost

- It knows that the generated results are bad, but does not know what results are good.

Teacher Forcing (Cont’d)

The author proposes feeding human generated responses to the generator for model updates.

- discriminator automatically assigns a reward of 1 to the human responses and feed it to the

generator to use this reward to update itself.

- Analogous to having a teacher intervene and force it to generate the true responses

Generator then updates itself using this reward on the human generated example only if the reward is

larger than the baseline value.

Pseudocode for the Algorithm

Result Again

Adversarially-Trained system generates higher-quality responses than previous baselines!

Notes

It did not show great performance on abstractive summarization task.

Maybe because adversarial training strategy is more beneficial to:

- Tasks in which there is a big discrepancy between the distributions of the generated sequences and

the reference target sequences

- Tasks in which input sequences do not bear all the information needed to generate the target

- in other words, there is no single correct target sequence in the semantic space.