1. Chatbot
Sequence to Sequence Learning
29 Mar 2017
Presented By:
Jin Zhang
Yang Zhou
Fred Qin
Liam Bui
Overview
Network
Architecture
Loss Function
Improvement
Techniques
3. Overview
Network
Architecture
Loss Function
Improvement
Techniques
LSTM for Language Model
• Language Model
Predict next word given the previous words
• RNN
Unable to learn long term dependency, not suitable for language model
• LSTM
3 sigmoid gates to control info flow
Understanding LSTM Networks: http://colah.github.io/posts/2015-08-Understanding-LSTMs/
4. Overview
Network
Architecture
Loss Function
Improvement
Techniques
• First step: which previous information to throw
away from the cell state
LSTM for Language Model
• Second step: what new information to be
stored in the cell state
- A sigmoid layer decides which values to
update
- A tanh layer creates new candidate values
C~t that could be added to the state
- Combine these two to create an update to
the state
• Third step: filter Ct and output only what we
want to output
Understanding LSTM Networks: http://colah.github.io/posts/2015-08-Understanding-LSTMs/
5. Seq2Seq model comprises of two language models:
• Encoder: a language model to encode input sequence into a fixed length vector (thought vector)
• Decoder: another language model to look at both thought vector and previous output to generate next words
Overview
Network
Architecture
Loss Function
Improvement
Techniques
Sequence To Sequence Model
Neural Machine Translation by Jointly Learning to Align and Translate: https://arxiv.org/abs/1409.0473
8. Overview
Network
Architecture
Loss Function
Improvement
Techniques
Generating a word is a multi-class classification task over all possible words, i.e. vocabulary.
W* = argmaxW P(W|Previous words)
Example :
I always order pizza with cheese and ……
mushrooms 0.15
pepperoni 0.12
anchovies 0.01
….
rice 0.0001
and 1e-100
Loss Function
9. Cross Entropy Loss:
Cross-Entropy:
Cross-Entropy for a sentence w1, w2, …, wn:
Overview
Network
Architecture
Loss Function
Improvement
Techniques
Evaluating Language Model: https://courses.engr.illinois.edu/cs498jh/Slides/Lecture04.pdf
Perplexity:
In practice, a variant called perplexity is usually used as metric to evaluate language models.
10. • Cross entropy can be seen as a measure of uncertainty
• Perplexity can be seen as “number of choices”
Overview
Network
Architecture
Loss Function
Improvement
Techniques
Cross entropy loss vs Perplexity:
• Entropy: ~2.58
• Perplexity: 6 choices
• Which statement do you prefer?
- The die has 6 faces
- The die has 2.58 entropy
• We can see perplexity as the average choices each time. The higher it is, the more “choices” of words
you have, then the more uncertain the language model is.
• Example: 6 faced balanced die. Each face is numbered from 1 to 6 so we have
11. Overview
Network
Architecture
Loss Function
Improvement
Techniques
Problem:
- The last state of the encoder contains mostly
information from the last elements of the
encoder sequence
- Inverse Input Sequence helps in some cases
How are you ?
I am fine .
Attention Mechanism:
- Allow each stage in decoder to look at any
encoder stages
- Decoder understand the input sentence more
and look at suitable positions to generate words
Neural Machine Translation by Jointly Learning to Align and Translate: https://arxiv.org/abs/1409.0473
12. Overview
Network
Architecture
Loss Function
Improvement
Techniques
Problem:
- The last state of the encoder contains mostly
information from the last elements of the
encoder sequence
- Inverse Input Sequence helps in some cases
Attention Mechanism:
- Allow each stage in decoder to look at any
encoder stages
- Decoder understand the input sentence more
and look at suitable positions to generate words
Neural Machine Translation by Jointly Learning to Align and Translate: https://arxiv.org/abs/1409.0473
Seq2Seq Seq2Seq with
attention
Sentence Length - 30 13.93 21.50
Sentence Length - 50 17.82 28.45
BLEU score on English-French Translation corpus
13. Overview
Network
Architecture
Loss Function
Improvement
Techniques
Problem:
- Maximizing conditional probabilities at each stage
might not lead to maximum full-joint probability.
- Storing all possible generated sentences are not
feasible due to resource limitation.
Possible output 2: Never been better
Possible output 1: I am fine
Beam Search:
- At each stage in decoder, store best M possible
outputs
Sequence to Sequence Learning: https://papers.nips.cc/paper/5346-sequence-to-sequence-learning-with-neural-networks.pdf
Conditional Probability: 0.6 0.4 1
Conditional Probability: 0.4 0.9 1
Full-joint
Probability:
0.24
0.36
Possible Output 1:
Possible Output 2:
Possible Output M:
How are you ?
I am fine .
…
14. Overview
Network
Architecture
Loss Function
Improvement
Techniques
Problem:
- Maximizing conditional probabilities at each stage
might not lead to maximum full-joint probability.
- Storing all possible generated sentences are not
feasible due to memory limitation.
Beam Search:
- At each stage in decoder, store best M possible
outputs
Sequence to Sequence Learning: https://papers.nips.cc/paper/5346-sequence-to-sequence-learning-with-neural-networks.pdf
Seq2Seq with
beam-size = 1
Seq2Seq with
beam size = 12
28.45 30.59
BLEU score on English-French Translation corpus.
Max sentence length 50
18. Overview
Network
Architecture
Loss Function
Improvement
Techniques
1. Reinforcement Learning:
Longer sentence is usually more interesting. So, we can use sentence length as rewards to further train the model:
• Action: Word choice
• State: Current generated sentence
• Reward: Sentence Length
2. Adversarial Training:
Make generated sentences look real using Adversarial training:
• Generative Model: generate sentences based on inputs
• Discriminant Model: tries to tell if a sentence is true response or generated response
• Objective: train generative model to “fool“ discriminant model
Adversarial Learning for Neural Dialogue Generation: https://arxiv.org/abs/1701.06547
Editor's Notes
First of all, let’s see a demo. This is a customer service chatbot demo. We can see that you can let it find an order as easy as chatting with a person. That’s why chatbot is a very hot topic. Many companies are working on various kinds of chatbots, including travel search engine, personal health companion and so on.
There are three ways to compare chatbots.
Retrieval-based models don’t generate any new text, they just pick a response from a predefined responses repository based on the input and context. In such case, retrieval-based methods don’t make grammatical mistakes. However, they may be unable to handle unseen cases for which no appropriate predefined response exists. For the same reasons, these models can’t refer back to contextual entity information like names mentioned earlier in the conversation.
However, Generative models don’t rely on pre-defined response. They generate new responses from scratch. Generative models are “smarter”. They can refer back to entities in the input and give the impression that you’re talking to a human. However, these models are hard to train, are quite likely to make grammatical mistakes (especially on longer sentences), and typically require huge amounts of training data. Generative models are “smarter”. They can refer back to entities in the input and give the impression that you’re talking to a human.
Chatbots can be built to support short-text conversations, such as FAQ chatbot, or long conversations, such as customer support chatbot.
Chatbots can be set to closed domain or open domain. The demo of this customer service chatbot is an example of the closed domain, in which the questions and answers are limited to specific area. In an open domain (harder) setting the user can take the conversation anywhere, such as siri.
Retrieval-based models (easier) use a repository of predefined responses and some kind of heuristic to pick an appropriate response based on the input and context. The heuristic could be as simple as a rule-based expression match, or as complex as an ensemble of Machine Learning classifiers. These systems don’t generate any new text, they just pick a response from a fixed set.
Due to the repository of handcrafted responses, retrieval-based methods don’t make grammatical mistakes. However, they may be unable to handle unseen cases for which no appropriate predefined response exists. For the same reasons, these models can’t refer back to contextual entity information like names mentioned earlier in the conversation. Generative models are “smarter”. They can refer back to entities in the input and give the impression that you’re talking to a human. However, these models are hard to train, are quite likely to make grammatical mistakes (especially on longer sentences), and typically require huge amounts of training data.
Short-Text Conversations (easier) where the goal is to create a single response to a single input. For example, you may receive a specific question from a user and reply with an appropriate answer. Then there are long conversations (harder) where you go through multiple turns and need to keep track of what has been said. Customer support conversations are typically long conversational threads with multiple questions.
In a closed domain (easier) setting the space of possible inputs and outputs is somewhat limited because the system is trying to achieve a very specific goal. Technical Customer Support or Shopping Assistants are examples of closed domain problems.
In an open domain (harder) setting the user can take the conversation anywhere. There isn’t necessarily have a well-defined goal or intention. The infinite number of topics and the fact that a certain amount of world knowledge is required to create reasonable responses makes this a hard problem.
The foundation of building a chatbot is language modelling. Generally speaking, a language model takes in a sequence of inputs, looks at each element of the sequence and tries to predict the next element of the sequence.
In theory, RNNs are absolutely capable of handling such “long-term dependencies.” A human could carefully pick parameters for them to solve toy problems of this form. Sadly, in practice, RNNs don’t seem to be able to learn them.
LSTMs are explicitly designed to avoid the long-term dependency problem.
The key to the LSTM is the cell state, easy for information to just flow along it unchanged.
The sigmoid layer outputs numbers between zero and one, describing how much of each component should be let through. A value of zero means “let nothing through,” while a value of one means “let everything through!”
An LSTM has three of these gates, to protect and control the cell state.
The first step in our LSTM is to decide what information we’re going to throw away from the cell state. This decision is made by a sigmoid layer called the “forget gate layer.” It looks at ht-1 and xt and outputs a number between 0 and 1 for each number in the cell state Ct-1. A
1 represents “completely keep this” while a 0 represents “completely get rid of this.”
Let’s go back to our example of a language model trying to predict the next word based on all the previous ones. In such a problem, the cell state might include the gender of the present subject, so that the correct pronouns can be used. When we see a new subject, we want to forget the gender of the old subject.
The next step is to decide what new information we’re going to store in the cell state. This has two parts. First, a sigmoid layer called the “input gate layer” decides which values we’ll update. Next, a tanh layer creates a vector of new candidate values, C~t, that could be added to the state. In the next step, we’ll combine these two to create an update to the state.
In the example of our language model, we’d want to add the gender of the new subject to the cell state, to replace the old one we’re forgetting.
It’s now time to update the old cell state, C~t-1into the new cell state C~t. The previous steps already decided what to do, we just need to actually do it.
We multiply the old state by ft, forgetting the things we decided to forget earlier. Then we add C~t*it .This is the new candidate values, scaled by how much we decided to update each state value.
In the case of the language model, this is where we’d actually drop the information about the old subject’s gender and add the new information, as we decided in the previous steps.
Finally, we need to decide what we’re going to output. This output will be based on our cell state, but will be a filtered version. First, we run a sigmoid layer which decides what parts of the cell state we’re going to output. Then, we put the cell state through tanh (to push the values to be between −1 and 1) and multiply it by the output of the sigmoid gate, so that we only output the parts we decided to.
For the language model example, since it just saw a subject, it might want to output information relevant to a verb, in case that’s what is coming next. For example, it might output whether the subject is singular or plural, so that we know what form a verb should be conjugated into if that’s what follows next.
RNNs can be used as language models for predicting future elements of a sequence given prior elements of the sequence. However, we are still missing the components necessary for building translation models since we can only operate on a single sequence, while translation operates on two sequences – the input sequence and the translated sequence.
Sequence to sequence models build on top of language models by adding an encoder step and a decoder step. In the encoder step, a model converts an input sequence into a thought vector. In the decoder step, a language model is trained on both the output sequence as well as the thought vector from the encoder. Since the decoder model sees an encoded representation of the input sequence as well as the output sequence, it can make more intelligent predictions about future words based on the current word.
For example, in a standard language model, we might see the word “crane” and not be sure if the next word should be about the bird or heavy machinery. However, if we also pass an encoder context, the decoder might realize that the input sequence was about construction, not flying animals. Given the context, the decoder can choose the appropriate next word and provide more accurate reply.
Now that we understand the basics of sequence to sequence modeling, we can consider how to build one. We will use LSTM as encoder and decoder.
The encoder takes a sequence(sentence) as input and processes one symbol(word) at each time step. Its objective is to convert a sequence of symbols into a fixed size feature vector that encodes only the important information in the sequence while losing the unnecessary information.
Each hidden state influences the next hidden state and the final hidden state can be seen as the summary of the sequence. This state is called the context or thought vector, as it represents the intention of the sequence. From the context, the decoder generates another sequence, one symbol(word) at a time. Here, at each time step, the decoder is influenced by the context and the previously generated symbols.
We can train the model using a gradient-based algorithm, update parameters of encoder and decoder, jointly maximize the log probability of the output sequence conditioned on the input sequence. Once the model is trained, we can make predictions
The context can be provided as the initial state of the decoder RNN or it can be connected to the hidden units at each time step. Now our objective is to jointly maximize the log probability of the output sequence conditioned on the input sequence.
whichever a model gives us the highest prop for all the words should be our model.
Evaluate per word perplexity. For the probability, we definitely think of cross entropy.
https://courses.engr.illinois.edu/cs498jh/Slides/Lecture04.pdf
by applying the chain rule, we can get perplexity per word.
Compressing an entire input sequence into a single fixed vector is challenging. The last state of the encoder contains mostly information from the last elements of the encoder sequence
This mechanism will hold onto all states from the encoder and give the decoder a weighted average of the encoder states for each element of the decoder sequence
During the decoding phase, we take the state of the decoder network, combine it with the encoder states, and pass this combination to a feedforward network. The feedforward network returns weights for each encoder state. We multiply the encoder inputs by these weights and then compute a weighted average of the encoder states.
BLEU (bilingual evaluation understudy): measures the correspondence between a machine's output and that of a human
BLEU = sum: max(word count in generated sentence, word count in referenced sentence)/total generated sentence length for each word
Maximizing conditional probabilities at each stage might not lead to maximum full-joint probability.
We could store all possible generated sentences so that we always find the maximum full-joint probability, but it would not be feasible.
A practical solution would be something in between.