Statistical Methods for NLP

Introduction, Text Mining, Linear Methods

of Regression
Sameer Maskey
Week 1, January 19, 2010
Course Information
! Course Website:
http://www.cs.columbia.edu/~smaskey/CS6998
! Discussions in courseworks
! Office hours Tues: 2 to 4pm, Speech Lab (7LW1), CEPSR
! Individual appointments in person or in phone can be set
by emailing the instructor : [email protected]
! Instructor: Dr. Sameer Maskey
! Prerequisites
" Probability, statistics, linear algebra, programming skill
" CS Account
Grading and Academic Integrity
! 3 Homework (15% each)
" Homework due dates are available in the class webpage
" You have 3 no penalty late days in total that can be used during
the semester
" Each additional late day (without approval) will be penalized, 20%
each day
! No midterm exam
! Final project (40%)
" It is meant for you to explore and do research NLP/ML topic of
your choice
" Project proposal due sometime in the first half of the semester
! Final Exam (15%)
! Collaboration allowed but presenting someone elses work
(including code) will result in automatic zero
Textbook
! For NLP topics we will use the following book:
" Speech and Language Processing (2
nd
Edition) by Daniel
Jurafsky and James H Martin
! For statistical methods/ML topics we will partly use
" Pattern Recognition and Machine Learning by Christopher
Bishop
" There are also two online textbooks which will be available
for the class, some readings may be assigned from these
" Other readings will be provided for the class online
Goal of the Class
! By the end of the semester
" You will have in-depth knowledge of several NLP and ML topics
and explore the relationship between them
" You should be able to implement many of the NLP/ML methods
on your own
" You will be able to frame many of the NLP problems in statistical
framework of your choice
" You will understand how to analytically read NLP/ML papers and
know the kind of questions to ask oneself when doing NLP/ML
research
Topics in NLP (HLT, ACL) Conference
! Morphology (including word segmentation)
! Part of speech tagging
! Syntax and parsing
! Grammar Engineering
! Word sense disambiguation
! Lexical semantics
! Mathematical Linguistics
! Textual entailment and paraphrasing
! Discourse and pragmatics
! Knowledge acquisition and representation
! Noisy data analysis
! Machine translation
! Multilingual language processing
! Language generation
! Summarization
! Question answering
! Information retrieval
! Information extraction
! Topic classification and information filtering
! Non-topical classification (sentiment/genre analysis)
! Topic clustering
! Text and speech mining
! Text classification
! Evaluation (e.g., intrinsic, extrinsic, user studies)
! Development of language resources
! Rich transcription (automatic annotation)
!
Topics in ML (ICML, NIPS) Conference
! Reinforcement Learning
! Online Learning
! Ranking
! Graphs and Embeddding
! Gaussian Processes
! Dynamical Systems
! Kernels
! Codebook and Dictionaries
! Clustering Algorithms
! Structured Learning
! Topic Models
! Transfer Learning
! Weak Supervision
! Learning Structures
! Sequential Stochastic Models
! Active Learning
! Support Vector Machines
! Boosting
! Learning Kernels
! Information Theory and Estimation
! Bayesian Analysis
! Regression Methods
! Inference Algorithms
! Analyzing Networks & Learning with Graphs
!
! Morphology (including word segmentation)
! Part of speech tagging
! Syntax and parsing
! Grammar Engineering
! Word sense disambiguation
! Lexical semantics
! Mathematical Linguistics
! Textual entailment and paraphrasing
! Discourse and pragmatics
! Knowledge acquisition and representation
! Noisy data analysis
! Machine translation
! Multilingual language processing
! Language generation
! Summarization
! Question answering
! Information retrieval
! Information extraction
! Topic classification and information filtering
! Non-topical classification (sentiment/genre analysis)
! Topic clustering
! Text and speech mining
! Text classification
! Evaluation (e.g., intrinsic, extrinsic, user studies)
! Development of language resources
! Rich transcription (automatic annotation)
!
! Reinforcement Learning
! Online Learning
! Ranking
! Graphs and Embeddding
! Gaussian Processes
! Dynamical Systems
! Kernels
! Codebook and Dictionaries
! Clustering Algorithms
! Structured Learning
! Topic Models
! Transfer Learning
! Weak Supervision
! Learning Structures
! Sequential Stochastic Models
! Active Learning
! Support Vector Machines
! Boosting
! Learning Kernels
! Information Theory and Estimation
! Bayesian Analysis
! Regression Methods
! Inference Algorithms
! Analyzing Networks & Learning with Graphs
!
NLP ML
Many Topics Related
Tasks #$ Solutions
Combine Relevant Topics
Topics We Will Cover in This Course
NLP - - ML
! Text Mining
! Text Categorization
! Information Extraction
! Syntax and Parsing
! Topic and Document Clustering
! Machine Translation
! Synchronous Chart Parsing
! Language Modeling
! Speech-to-Speech Translation
! Evaluation Techniques
Linear Models of Regression
Linear Methods of Classification
Support Vector Machines
Kernel Methods
Hidden Markov Model
Maximum Entropy Models
Conditional Random Fields
K-means, KNN
Expectation Maximization
Spectral Clustering
Viterbi Search, Beam Search
Graphical Models
Belief Propogation
Text Mining
! Data Mining: finding nontrivial patterns in databases
that may be previously unknown and could be useful
! Text Mining:
" Find interesting patterns/information from unstructured text
" Discover new knowledge from these patterns/information
! Information Extraction, Summarization, Opinion
Analysis, etc can be thought as some form of text
mining
! Let us look at an example
Patterns in Unstructured Text
Patterns may exist
in unstructured text
Some of these patterns
could be exploited to
discover knowledge
All Amazon reviewers may not
rate the product, may just
write reviews, we may have to
infer the rating based on text review
Review of a camera in Amazon
Text to Knowledge
! Text
" Words, Reviews, News Stories, Sentences,
Corpus, Text Databases, Real-time text, Books
! Knowledge
" Ratings, Significance, Patterns, Scores, Relations
Many methods to use
for discovering
knowledge from text
Unstructured Text $ Score
Facebooks Gross National Happiness Index
! Facebook users update their status
" is writing a paper
" has flu %
" is happy, yankees won!
! Facebook updates are unstructured text
! Scientists collected all updates and analyzed them
to predict Gross National Happiness Index
Facebooks Gross National Happiness Index
How do you think they extracted this SCORE from a
TEXT collection of status updates?
Facebook Blog Explains
! The result was an index that measures how happy
people on Facebook are from day-to-day by looking
at the number of positive and negative words they're
using when updating their status. When people in
their status updates use more positive words - or
fewer negative words - then that day as a whole is
counted as happier than usual.
Looks like they are COUNTING!
+ve and ve words in status updates
Lets Build Our NLP/ML Model to Predict
Happiness &
! Simple Happiness Score
" Our simpler version of happiness index compared to
facebook
" Score ranges from 0 to 10
! There are a few things we need to consider
" We are using status updates words
" We do not know what words are positive and
negative
" We do not have any training data
Our Prediction Problem
! Training data
" Assume we have N=100,000 status updates
" Assume we have a simple list of positive and negative words
" Let us also assume we asked a human annotator to read each of
the 100,000 status update and give a happiness Score (Y
i
)
between 0 to 10
! is writing a paper (Y
1
= 4)
! has flu % (Y
2
= 1.8)
! .
! .
! .
! is happy, game was good! (Y
100,000
= 8.9)
! Test data
" likes the weather (Y
100,001
= ? )
Given labeled set of 100K
Status updates, how do we build
Statistical/ML model that
will predict the score for a
new status update
! What kind of feature can we come up with that would
relate well with happiness score
! How about represent status update as
" Count (+ve words in the sentence) (not the ideal
representation, will better representation letter)
" For the 100,000
th
sentence in our previous example:
! is happy, game was good. Count is 2
! Status Update 100,000
th
is represented by
" (X
100000
= 2, Y
100000
= 8.9)
Representing Text of Status Updates As a
Vector
Modeling Technique
! We want to predict happiness score (Y
i
) for a new status
update
! If we can model our training data with a statistical/ML model,
we can do such prediction
" (1, 4)
" (0, 1.8)
" .
" .
" .
" (2, 8.9)
! What modeling technique can we use?
" Linear Regression is one choice
Xi Yi
,
Linear Regression
$ We want to find a function that given our x it would
map it to y
$ One such function :
$ Different values of thetas give different functions
$ What is the best theta such that we have a
function that makes least error on predictions when
compared with y
Predicted vs. True
Sum of Squared Errors
! Plugging in f(x) and averaging the error across all training
data points we get the empirical loss
f(x
i
)
y
i
x
i

y
Finding the Minimum
$ We can (but not always) find a minimum of a
function by setting the derivative or partial derivatives
to zero
$ Here we can take partials on thetas and set them to
zero
Solving for Weights
Empirical Loss is Minimized With Given
Values for the Parameters
$ Solving the previous equations we get following values
for the thetas
Implementing Simple Linear Regression
! Given our training data on status update with happiness score
" (1, 4)
" (0, 1.8)
" .
" .
" .
" (2, 8.9)
Xi Yi
,
Training Our Regression Model:
Just need to implement for loop
that computes numerators and
denominators in equations here.
And we get optimal thetas
For Prediction/Testing:
Given optimal thetas, plug in the
x value in our equation to get y
Simple Happiness Scoring Model too
Simple?
! So far we have a regression model that was
trained on a training data of facebook status
updates (text) and labeled happiness score
! Status updates words were mapped to one
feature
" Feature counted number of +ve words
! Maybe too simple?
" How can we improve the model?
" Can we add more features?
! How about count of ve words as well
Let Us Add One More Feature
! Adding one more feature Z
i
representing count of ve words, now
training data will look like the following
" (1, 3, 4)
" (0, 6,1.8)
" .
" .
" .
" (2, 0, 8.9)
! What would our linear regression
function would look like
Xi Zi
,
Yi
,
Estimation of y i.e. f(x,z) is now a plane instead of a line
[3]
Regression Function in Matrix Form
! Remember our regression function in 2D looked
like
! Representing in Matrix form we get,
! And empirical loss will be
Adding Features
! In K dimensions the regression function f(x) we estimate will
look like
! So the empirical loss would
! Representing with matrices
Empirical Loss with K Features and N
Data Points in Matrix Representation
! Representing empirical loss in Matrix form
Y X !
Solve by Setting Partial Derivatives to
Zero
! Remember, to find the minimum empirical loss we set the
partial derivatives to zero
! We can still do the same in matrix form, we have to set the
derivatives to zero
! Solving the above equation we get our best set of parameters
Implementation of Multiple Linear
Regression
! Given out N training data points we can build X and Y matrix
and perform the matrix operations
! Can use MATLAB
! Or write your own, Matrix multiplication implementation
! Get the theta matrix
! For any new test data plug in the x values (features) in our
regression function with the best theta values we have
Back to Our Happiness Prediction
Regression Model
! Xi1 represented count of +ve words
! (Xi1, Yi) pair were used to build simple linear regression model
! We added one more feature Xi2, representing count of ve words
! (Xi1, Xi2, Yi) can be used to build multiple linear regression model
! Our training data would look like
" (1, 3, 4)
" (0, 6,1.8)
" .
" .
" .
" (2, 0, 8.9)
! From this we can build X and Y Matrix and find the best theta values
! For N Data points, we will get Nx3 X matrix, Nx1 Y matrix and 3X1 ! matrix
Xi1 Xi2
,
Yi
,
More Features? Feature Engineering
! So far we have only two features, is it good enough?
! Should we add more features?
! What kind of features can we add?
" Ratio of +ve/-ve words
" Normalized count of +ve words
" Is there a verb in the sentence?
! We need to think what are the kinds of information that may
better estimate the Y values
! If we add above 3 features, what is the value of K?
Polynomial Regression
Polynomial Regression
K Features, M Order Polynomial and N
Data Points
! With K=1, we get a regression line, with K=2 we get
a plane
! With M=1 we get a straight line or plane
! With M=2 we get a curved line or plane
! So with K=2 and M=2 ?
Trend Surface
Trend Surfaces for different orders of polynomial [1]
Overfitting
! Higher order of polynomial should be used with caution though
! Higher order polynomial can fit the training data too closely
especially when few training points, with the generalization error
being high
! Leave one out cross validation allows to estimate generalization
error better
" If N data points use N-1 data points to train and use 1 to test
Higher order of polynomial overfitting with few data points [2]
Testing Our Model
! Our goal was to build the best statistical model that
would automate the process of scoring a chunk of
text (Happiness Score)
! How can we tell how good is our model?
! Remember previously we said let us assume we
have 100,000 status updates
! Instead of using all 100K sentences let use the first
90K to build the model
! Use rest of 10K to test the model
10-fold Cross Validation
! 10 fold cross validation
" We trained on first 90K (1 to 90,000)
" Tested on (90,001 to 100,000)
" But we can do this 10 times if we select different 10K of test data
point each time
10k 10k 10k 10k 10k 10k 10k 10k 10k 10k
10k 10k 10k 10k 10k 10k 10k 10k 10k 10k
10k 10k 10k 10k 10k 10k 10k 10k 10k 10k

! 10 experiments, build model and test times with 10 different sets of
training and test data
! Average the accuracy across 10 experiments
! We can do any N-fold cross validation to test our model
Exp1
Exp2
Exp10
Scores from Text, What Else Can They
Represent?
! Given a facebook status update we can predict
happiness score
! But we can use the same modeling technique in
many other problems
" Summarization: Score may represent importance
" Question Answering: Score may represent relevance
" Information extraction : Score may represent relation
! We need to engineer features according to the
problem
! Many uses of the statistical technique we learned
today
Reviews to Automatic Ratings
Model
Rating
Features
Scores
Y X
Statistical Model
Features
X
TRAIN
PREDICT
Unstructured Text to Binary Labels
! Let us change the type of problem a bit
! Instead of a real valued happiness score between 0 and 10,
let us assume our annotators just provide unhappy (0) or
happy (1)
" Or it can be Amazon review for a product dislike (0) and like (1)
! Can we and should we still model this kind data with
regression?
Gaussian Noise
Class Prediction from Text
! If y outputs are binary classes we may want
to use a different modeling technique
! Binary classifiers could model such data
! We need to chose our models according to
the problem we are handling
! We probably need better representation of
text as well
Readings
! 1.1, 3.1, 4.1 Bishop Book
! 23.1.1, 23.1.2, 23.1.3 Jurafsky & Martin Book
References
! [1] http://biol09.biol.umontreal.ca/PLcourses/Polynomial_regression.pdf
! [2] Christopher Bishop, Pattern Recognition and Machine Learning, Springer, 2006
! [3] Hastie, Tibshirani and Friedman, Elements of Statistical Learning, 2001