NLP and Word Embeddings - web.stanford.edu · NLP and Word Embeddings GloVe word vectors. Andrew Ng...
Transcript of NLP and Word Embeddings - web.stanford.edu · NLP and Word Embeddings GloVe word vectors. Andrew Ng...
deeplearning.ai
NLPandWordEmbeddings
Wordrepresentation
AndrewNg
Word representationV = [a, aaron, …, zulu, <UNK>]
1-hot representation
Apple(456)
Orange(6257) I want a glass of orange ______.
I want a glass of apple______.
King(4914)000⋮1⋮000
Woman(9853)00000⋮1⋮0
Man(5391)
0000⋮1⋮00
Queen(7157)
00000⋮1⋮0
0⋮1⋮00000
00000⋮1⋮0
AndrewNg
Featurized representation: word embeddingApple(456)
Orange(6257)
King(4914)
Woman(9853)
Man(5391)
Queen(7157)
I want a glass of orange ______.I want a glass of apple______.
-0.95 0.97 0.00 0.01
0.93 0.95 -0.01 0.00
0.7 0.69 0.03 -0.02
0.02 0.01 0.95 0.97
AndrewNg
Visualizing word embeddings
fish
dogcat
applegrape
orangeonethree
two
four
king
man
queen
woman
[van der Maaten and Hinton., 2008. Visualizing data using t-SNE]
deeplearning.ai
NLPandWordEmbeddings
Usingwordembeddings
AndrewNg
Named entity recognition example
Sally Johnson is an orange farmer
1 1 0 0 0 0
Robert Lin is an apple farmer
AndrewNg
Transfer learning and word embeddings
1. Learn word embeddings from large text corpus. (1-100B words)
(Or download pre-trained embedding online.)
2. Transfer embedding to new task with smaller training set. (say, 100k words)
3. Optional: Continue to finetune the word embeddings with newdata.
AndrewNg
Relation to face encoding
⋮
$(&)
⋮
⋮
$(()
)*
[Taigman et. al., 2014. DeepFace: Closing the gap to human level performance]
f($(&))
f($(())
deeplearning.ai
NLPandWordEmbeddings
Propertiesofwordembeddings
AndrewNg
AnalogiesApple(456)
Orange(6257)
King(4914)
Woman(9853)
Man(5391)
Queen(7157)
Gender
Royal
Age
Food
−10.010.030.09
10.020.020.01
-0.950.930.700.02
0.970.950.690.01
0.00-0.010.030.95
0.010.00-0.020.97
[Mikolov et. al., 2013, Linguistic regularities in continuous space word representations]
AndrewNg
Analogies using word vectorsfish
dog
cat
applegrape
orangeone
three
two
four
kingman
queen
woman
()*+ − (,-)*+ ≈ (/0+1 − (?
AndrewNg
Cosine similarity
345((,, (/0+1 − ()*+ + (,-)*+)
Man:Woman as Boy:GirlOttawa:Canada as Nairobi:KenyaBig:Bigger as Tall:TallerYen:Japan as Ruble:Russia
deeplearning.ai
NLPandWordEmbeddings
Embeddingmatrix
AndrewNg
Embedding matrix
In practice, use specialized function to look up an embedding.
deeplearning.ai
NLPandWordEmbeddings
Learningwordembeddings
AndrewNg
Neural language modelI want a glass of orange ______.4343 9665 1 3852 6163 6257
I
want
a
glass
of
orange
*+,+,
*-../*0*,1/2*.0.,*.2/3
4
44
44
4
5+,+,
5-../505,1/25.0.,
5.2/3[Bengio et. al., 2003, A neural probabilistic language model]
AndrewNg
Other context/target pairsI want a glass of orange juice to go along with my cereal.
Context: Last 4 words.
4 words on left & right
Last 1 word
Nearby 1 word
deeplearning.ai
NLPandWordEmbeddings
Word2Vec
AndrewNg
Skip-gramsI want a glass of orange juice to go along with my cereal.
[Mikolov et. al., 2013. Efficient estimation of word representations in vector space.]
AndrewNg
ModelVocab size = 10,000k
AndrewNg
Problems with softmax classification
! " # = %&'()*∑ %&,()*-.,...01-
How to sample the context #?
deeplearning.ai
NLPandWordEmbeddings
Negativesampling
AndrewNg
Defining a new learning problem
I want a glass of orange juice to go along with my cereal.
[Mikolov et. al., 2013. Distributed representation of words and phrases and their compositionality]
AndrewNg
Model
Softmax: ! " # = %&'()*∑ %&,()*-.,...01-
context wordorangeorangeorange
juicekingbook
target?
theof
orangeorange
100
00
AndrewNg
Selecting negative examples
context wordorangeorangeorange
juicekingbook
target?
theof
orangeorange
100
00
deeplearning.ai
NLPandWordEmbeddings
GloVe wordvectors
AndrewNg
GloVe (global vectors for word representation)
I want a glass of orange juice to go along with my cereal.
[Pennington et. al., 2014. GloVe: Global vectors for word representation]
AndrewNg
Model
AndrewNg
A note on the featurization view of word embeddings
minimize ∑ ∑ ( )*+ ,*-.+ + 0* − 0+2 − log)*+678,888
+:778,888*:7
King(4914)
Woman(9853)
Man(5391)
Queen(7157)
-0.950.930.700.02
0.970.950.690.01
−10.010.030.09
10.020.020.01
GenderRoyalAgeFood
deeplearning.ai
NLPandWordEmbeddings
Sentimentclassification
AndrewNg
Sentiment classification problem! "
The dessert is excellent.
Service was quite slow.
Good for a quick meal, but nothing special.
Completely lacking in good taste, good service, and good ambience.
AndrewNg
Simple sentiment classification model
The
desert
is
excellent
#$%&$
#&'($
#'(%'
#)*$+
,
,
,
,
-$%&$
-&'($
-'(%'
-)*$+
8928 2468 4694 3180The dessert is excellent
“Completely lacking in good taste, good service, and good ambience.”
AndrewNg
RNN for sentiment classification
Completely lacking in good …. ambience
, , , , ,
-*$6& -'%(( -''&7 -)$$& -))+
"8
softmax
⋯:;+< :;*< :;&< :;)< :;'< :;*+<
deeplearning.ai
NLPandWordEmbeddings
Debiasingwordembeddings
AndrewNg
The problem of bias in word embeddings
[Bolukbasi et. al., 2016. Man is to computer programmer as woman is to homemaker? Debiasing word embeddings]
Man:Woman as King:Queen
Man:Computer_Programmer as Woman:
Father:Doctor as Mother:
Word embeddings can reflect gender, ethnicity, age, sexual orientation, and other biases of the text used to train the model.
Homemaker
Nurse
AndrewNg
Addressing bias in word embeddings1. Identify bias direction.
2. Neutralize: For every word that is not definitional, project to get rid of bias.
3. Equalize pairs.
[Bolukbasi et. al., 2016. Man is to computer programmer as woman is to homemaker? Debiasing word embeddings]