Presentation 20 Dec 2015 Dialogue systems

реклама
Dialogue Systems
End-to-End
Review
Vladislav Belyaev
DeepHackLab
20 December 2015
Agenda
1. Dialogue System – objectives and problems
2. Architectures
2.1. Char-rnn
2.2. Sequence-to-sequence (seq2seq)
2.3. Hierarchical recurrent encoder-decoder (HRED)
2.4. Attention with Intention (AWI)
2.5. Memory Networks
3. Summary
Dialogue System
A.M. Turing. Computing machinery and intelligence. Mind, pages 433–460, 1950
Dialogue System
Goal driven and Non-goal driven
Deterministic (rule based, IR)
and Generative
Large dialogue corpora (0.1 or 1 billions of words)
Natural and Unnatural + can depend on external unobserved events
Preprocessing
Segmentation
Evaluation metric/Problems
Remove anomalies
(acronyms, slang, misspellings
and phonemicization)
Tokenization, stemming
and lemmatization
Speaker segmentation
(diarisation)
How to measure sense?
Conversation segmentation
Serban J.V. et al. (2015), A Survey of Available Corpora For Building Data-Driven Dialogue Systems
Generalization to previously
unseen situations
Highly generic responses
Diversity of corpora
Dialogue System
Evaluation
metrics
Goal-related performance criteria
(including user simulator)
Non-goal
“Naturalness” – human evaluation (+)
BLUE/METEOR score (-)
Next utterance classification (+)
Word perplexity (+)
Response diversity (in comb)
A.M. Turing. Computing machinery and intelligence. Mind, pages 433–460, 1950
Serban J.V. et al. (2015), A Survey of Available Corpora For Building Data-Driven Dialogue Systems
Dialogue System - Business
Without neural nets!
http://www.nextit.com/case-studies/amtrak/
http://www.nextit.com/case-studies/alaska-airlines/
Dialogue System - Business
Development costs
(hand-crafted features)
РАЗРАБОТКА НА ОСНОВЕ ЛИНГВИСТИЧЕСКИХ ПРАВИЛ
Начало разработки системы
Обучение
нейросети
Высокая длительность разработки (3-5 месяцев)
Ввод в эксплуатацию
+
Domain limitations
Высокая стоимость
Создание базы знаний
Тестирование
Ввод в эксплуатацию
Быстрая разработка
(2-3 недели)
Низкая
стоимость
Тестирование
НЕЙРОСЕТЕВАЯ
Architectures
Opportunities
Data Size
Papers
Code
Char-rnn
Suits well for morphologically
rich languages
Memory limited by RAM +
Hard to remember facts
Around several 10Ms of words
is better
http://karpathy.github.io/2015/
05/21/rnn-effectiveness/
http://arxiv.org/abs/1508.06615
https://github.com/karpathy/charrnn
https://github.com/yoonkim/lstmchar-cnn
Seq2seq
Don’t contain dialogue states
No really long-term
dependencies
(tricky play)
For NCM
Help desk – 30M of tokens
OpenSubtitles – 923M of tokens
http://arxiv.org/abs/1506.05869
https://www.tensorflow.org/versions
/master/tutorials/seq2seq/index.htm
l#sequence-to-sequence-models
https://github.com/macournoyer/ne
uralconvo
HRED
Has a dialogue state
representation
Hard to remember facts
>1B of tokens recommended
http://arxiv.org/abs/1507.04808
http://arxiv.org/abs/1507.02221
https://github.com/julianser/hed-dlg
https://github.com/julianser/hed-dlgtruncated
AWI
Amazing results (a dialogue
state and attention), but no
code and in-house dataset
4.5M of tokens (1000 dialogues)
http://arxiv.org/abs/1510.08565
-
MemNN
Can perform QA,
recommendation and chitchat
http://fb.ai/babi
3.5M training examples (4 ds for
4 tasks)
http://arxiv.org/abs/1511.06931
https://github.com/carpedm20/Mem
N2N-tensorflow
https://github.com/facebook/SCRNN
s
Char-rnn
http://karpathy.github.io/2015/05/21/rnn-effectiveness/
Y. Kim et al. (2015), Character-Aware Neural Language Models
Sequence-to-sequence
4 LSTM layers with
1000 memory cells (1 layer = +10% perp)
Input vocabulary 160k words
Output vocab 80k words
1000 for word embeddings (for 300M words)
Grouping sentences by length (5.8 perp)
Reversing sentences (4.7 perp)
Ilya Sutskever, Oriol Vinyals, Quoc V. Le (2014), Sequence to Sequence Learning with Neural Networks
Sequence-to-sequence
Ilya Sutskever, Oriol Vinyals, Quoc V. Le (2014), Sequence to Sequence Learning with Neural Networks
https://www.tensorflow.org/versions/master/tutorials/seq2seq/index.html#sequence-to-sequence-models
http://googleresearch.blogspot.ca/2015/11/computer-respond-to-this-email.html
Sequence-to-sequence
A single layer LSTM with
1024 memory cells
The vocabulary consists of the
most common 20K words
30M of tokens
Model achieved a perplexity of 8,
An n-grammodel achieved 18
2 layers LSTM with
4096 memory cells + 2048 linear units
The vocabulary consists of the most
common frequent 100K words
923M of tokens
Model achieved a perplexity of 17,
An 5-grammodel achieved 28
Oriol Vinyals, Quoc V. Le (2015), A Neural Conversational Model
Hierarchical recurrent encoder-decoder
Iulian V. Serban, et al. (2015), Building End-To-End Dialogue Systems Using Generative Hierarchical Neural Network Models
Hierarchical recurrent encoder-decoder
For the baseline RNN, we tested hidden
state spaces dh = 200, 300 and 400. For
HRED we experimented with encoder and
decoder hidden state spaces of size 200,
300 and 400. Based on preliminary results
and due to GPU memory limitations, we
limited ourselves to size 300 when not
bootstrapping
or
bootstrapping
from
Word2Vec, and to size 400 when
bootstrapping from SubTle. Preliminary
experiments showed that the context
RNN state space at and above 300
performed similarly, so we fixed it at 300
when not bootstrapping or bootstrapping
from Word2Vec, and to 1200 when
bootstrapping from SubTle. For all models,
we used word embedding of size
400 when bootstrapping from SubTle and of
size 300 otherwise
Iulian V. Serban, et al. (2015), Building End-To-End Dialogue Systems Using Generative Hierarchical Neural Network Models
Hierarchical recurrent encoder-decoder
# Gradients will be truncated after 80 steps. This seems like a fair start.
state['max_grad_steps'] = 80
2015-12-20 11:37:04,223: search: DEBUG: adding sentence [16, 306, 9294, 17, 15421, 10, 1] from beam 0
state['qdim_encoder'] = 3000
2015-12-20 11:37:04,223: search: DEBUG: partial -> боже , нет . нет , мама
state['qdim_decoder'] = 3000
2015-12-20 11:37:04,223: search: DEBUG: partial -> - вот гений ! - рад это
# Dimensionality of dialogue hidden layer
2015-12-20 11:37:04,223: search: DEBUG: partial -> конечно , он твой приятель один .
2015-12-20 11:37:04,223: search: INFO: RandomSampler : sampling step 7, beams alive 3
state['sdim'] = 1000
2015-12-20 11:37:04,330: search: DEBUG: partial -> боже , нет . нет , мама .
# Dimensionality of low-rank approximation
2015-12-20 11:37:04,330: search: DEBUG: partial -> - вот гений ! - рад это слышать
2015-12-20 11:37:04,330: search: DEBUG: partial -> конечно , он твой приятель один . он
state['rankdim'] = 1000
30k vocab, 77M words, perp 39.0644
2015-12-20 11:37:04,330: search: INFO: RandomSampler : sampling step 8, beams alive 3
2015-12-20 11:37:04,437: search: DEBUG: adding sentence [212, 11, 35, 10, 35, 11, 211, 10, 1] from beam 0
2015-12-20 11:37:04,437: search: DEBUG: partial -> - вот гений ! - рад это слышать .
2015-12-20 11:37:04,437: search: DEBUG: partial -> конечно , он твой приятель один . он тоже
2015-12-20 11:37:04,437: search: INFO: RandomSampler : sampling step 9, beams alive 2
2015-12-20 11:37:04,534: search: DEBUG: adding sentence [18, 77, 3610, 22, 18, 458, 20, 1217, 10, 1] from beam 0
2015-12-20 11:37:04,535: search: DEBUG: partial -> конечно , он твой приятель один . он тоже там
2015-12-20 11:37:04,535: search: INFO: RandomSampler : sampling step 10, beams alive 1
2015-12-20 11:37:04,624: search: DEBUG: partial -> конечно , он твой приятель один . он тоже там был
2015-12-20 11:37:04,624: search: INFO: RandomSampler : sampling step 11, beams alive 1
2015-12-20 11:37:04,712: search: DEBUG: partial -> конечно , он твой приятель один . он тоже там был .
2015-12-20 11:37:04,713: search: INFO: RandomSampler : sampling step 12, beams alive 1
2015-12-20 11:37:04,801: search: DEBUG: adding sentence [139, 11, 25, 162, 846, 151, 10, 25, 119, 98, 62, 10, 1] from
beam 0
2.68409920732: боже , нет . нет , мама .
3.10118466318: - вот гений ! - рад это слышать .
3.28937990442: и убил тебя .
3.72474860085: конечно , он твой приятель один . он тоже там был .
5.60221603406: ты дом налоги и болтаешь .
Hierarchical recurrent encoder-decoder
# Gradients will be truncated after 80 steps. This seems like a fair start.
state['max_grad_steps'] = 80
User – почему ты думаешь, что ИИ безопасен?
state['qdim_encoder'] = 3000
state['qdim_decoder'] = 3000
AI - потому что это правда .
# Dimensionality of dialogue hidden layer
User – а кто еще знает об этом?
state['sdim'] = 1000
AI - она не в курсе .
# Dimensionality of low-rank approximation
User – кто она?
state['rankdim'] = 1000
30k vocab, 77M words, perp 39.0644
AI - я - ваша дочь .
User – у меня нет детей!
AI - ладно .
User – и вообще, надо бы отключить питание за такие ответы.
AI - боже , нет . нет , мама .
User – ок, значит ИИ безопасен, да?
AI - как я и говорил ...
User – а чем ты еще занят?
AI - множество забот , ваша честь .
User – погоди ка, а куда делась дочь?
AI - нет , он просто хочет , чтобы ей стало лучше .
User – кто он? сколько вас там?
AI - но она здесь .
Attention with Intention
4.5M of tokens (1000 dialogues)
Kaisheng Yao et al. (2015), Attention with Intention for a Neural Network Conversation Model
AWI
Kaisheng Yao et al. (2015), Attention with Intention for a Neural Network Conversation Model
Memory Networks
With additional modifications to
construct both long-term and shortterm context memories
Sukhbaatar et al. (2015), End-To-End Memory Networks
J. Dodge et al. (2015), Evaluating prerequisite qualities for learning end-to-end dialog systems
Memory Networks
J. Dodge et al. (2015), Evaluating prerequisite qualities for learning end-to-end dialog systems
Memory Networks
J. Dodge et al. (2015), Evaluating prerequisite qualities for learning end-to-end dialog systems
Summary
Our vision (DeepHackLab + NN&DL LAB MIPT)
Large corpora
(natural and
unnatural)
Pre-processing
module
Where to find?
Russian or not?
Russian… 
How many words
are enough?
Do we need words?
What is the best
representation?
End-to-end
module
Architecture?
Natural language generation
The best suitable architecture?
Dialogue policy learning
Reinforcement learning?
Dialogue state tracking
What is the best representation?
Q&A
Knowledge
module
Architecture?
How to put all together?
Evaluation
module
How to measure sense?
How much is enough?
T. Mikolov et al. (2015), A Roadmap towards Machine Intelligence
J. Schmidhuber (2015), On Learning to Think: Algorithmic Information Theory for Novel
Combinations of Reinforcement Learning Controllers and Recurrent Neural World Models
http://qa.deephack.me/
Thank you for attention!
Questions?
Bibliography
Key refs
1. Serban J.V. et al. (2015), A Survey of Available Corpora For Building Data-Driven Dialogue Systems
http://arxiv.org/abs/1512.05742
2. Y. Kim et al. (2015), Character-Aware Neural Language Models
http://arxiv.org/abs/1508.06615
3. Ilya Sutskever, Oriol Vinyals, Quoc V. Le (2014), Sequence to Sequence Learning with Neural Networks
http://arxiv.org/abs/1409.3215
4. Oriol Vinyals, Quoc V. Le (2015), A Neural Conversational Model
http://arxiv.org/abs/1506.05869
5. Iulian V. Serban, et al. (2015), Building End-To-End Dialogue Systems Using Generative Hierarchical Neural Network Models
http://arxiv.org/abs/1507.04808
6. Kaisheng Yao et al. (2015), Attention with Intention for a Neural Network Conversation Model
http://arxiv.org/abs/1511.06931
7. J. Dodge et al. (2015), Evaluating prerequisite qualities for learning end-to-end dialog systems
http://arxiv.org/abs/1511.06931
8. T. Mikolov et al. (2015), A Roadmap towards Machine Intelligence
http://arxiv.org/abs/1511.08130
9. J. Schmidhuber (2015), On Learning to Think: Algorithmic Information Theory for Novel Combinations of Reinforcement Learning Controllers and Recurrent Neural World
Models
http://arxiv.org/abs/1511.09249
Скачать