The document discusses building message bots using neural networks. It provides an overview of recurrent neural networks (RNNs) including long short-term memory networks (LSTMs) and gated recurrent units (GRUs). Applications mentioned include machine translation, question answering, conversations, image captioning, and troubleshooting chatbots. Sequence-to-sequence models with attention and memory networks are also summarized as approaches for natural language processing tasks.
2. Register New Bot
• The callback URL must use HTTPS
• Channel ID, Secret and MID
• Server IP Whitelist
• Dummy friend for testing.
• Trial bot can send/receive
message.
• Business account support Mission
Sticker, Link Message, Make
browsing your website from the
LINE app seamless for users.
3. Mission Sticker Example :
Tokyo Disney
http://official-blog.line.me/tw-biz/archives/57396837.html
6. Message Bot
• ELIZA(1966, http://nlp-addiction.com/eliza/)
• seemed to be able to fool users into believing that they
were conversing with a real human
• recognition of cue words or phrases in the input, and the
output of corresponding pre-prepared or pre-programmed
responses
29. Sequence to sequence
Neural Network
* On the WMT’14 English to French translation task,we obtained a
BLEU score of 34.81
* This is by far the best result achieved by direct translation with large
neural net-works. For comparison, the BLEU score of an SMT
baseline on this dataset is 33.30
venturebeat.com
“Google Translate could become more accurate soon thanks to deep
learning” 2016/03
engadget.com
“Google is using neural network to improve Translate” 2016/03
30. Sequence to sequence
Neural Network
* Each box in the picture most commonly a GRU cell or an LSTM cell.
RNN
x
RNN RNN RNN
time
RNN
y
RNN
EOF
encoder decoder
Good
RNN
Job
context
vector
31. Sequence to sequence
Neural Network
• Reversing source words improve test BLEU scores
time
EOF
Good Job
context
vector
layer 1
layer 3
layer 4
32M 32M
• Made sure that all sentences in a minibatch are roughly of the same length
32. Sequence to sequence
Neural Network
• We can use this model for machine translation, question/
answering, and conversations without major changes in
the architecture
• A potential issue with this encoder–decoder approach is
that a neural network needs to be able to compress all the
necessary information of a source sentence into a fixed-
length vector. This may make it difficult for the neural
network to cope with long sentences,
33. Application
. IT Helpdesk
Troubleshooting dataset
Google Inbox App
Smart Reply
announce at 2015/11
Machine:hi
Human:hello
Machine:could you please let me know
what are the operating systems you are
using ?
Human:i am using linux
Machine:ok it was in the ssl vpn on which
you stated it was broken
….
Image Recognition
announce at 2015/07, Not Seq2Seq
34.
35. Sequence to sequence Neural
Network+Attention(RNNSearch)
RNN
x0
RNN
x1
RNN
x2
h0
f h1
f
RNN
x2
RNN
x1
RNN
x0
h0
bh1
b
time
RNN
EOF
Good
RNN
Job
[h1
b,h1
f]
[h0
b,h0
f]
c0 c1
+ +
a00
a01
a10
a11
context
vector
annotation
bi-directional RNN
37. End-to-End Memory Network :
Single Hop
input story
weighted sum{c1,c2,c3}
{a1,a2,a3}
softmax
{0.2, 0.1, 0.7}
dot product
input question
WC
WA
?
u1
WB
+o1 u2
softmax
WO
{ :0.9, :0.1}
constructing memory vectors with Bag-of-Words
m1=V +V +V +V
Memory Module Controller
38. weighted sum
End-to-End Memory Network :
Two Hops
input story
softmax
dot product
input question
WC2
WA1
?
u1
WB
+o2 u3
softmax
WO
{ :0.9, :0.1}
Memory Module Controller
+o1
u2
weighted sum
softmax
dot product
weight sharing methods
1. Ak+1=Ck,
2. A1=A2=…=AK and C1=C2=…=CK
WA2
WC1
http://cims.nyu.edu/~sainbar/memnn_nips_pdf.pdf
39. Memory Network
Compare with RNNSearch, memory network
can be considered as
an attention model with multiple hops(lookup) and out of order
40.
41. Ask Me Anything: Dynamic Memory
Networks for Natural Language
Processing
http://cs224d.stanford.edu/lectures/CS224d-Lecture17.pdf
42. Reference
Sequence to Sequence Learning with Neural Networks.
Ilya Sutskever(OpenAI Research Director, Andrew Ng PostDoc), Oriol Vinyals &
Quoc V. Le
2014/09
Neural Machine Translation by Jointly Learning to Align and Translate.
Dzmitry Bahdanau(Theano Contributor), Kyunghyun Cho &Yoshua Bengio
2014/09
A Neural Conversational Model
Oriol Vinyals(Tensorflow Contributor) & Quoc Le
2015/01
Memory Networks
Jason Weston, Sumit Chopra & Antoine Bordes
2014/10
End-To-End Memory Networks
Sainbayar Sukhbaatar, Arthur Szlam, Jason Weston & Rob Fergus
2015/01