Transcript Improving Text Classification Accuracy by Augmenting

Automatic Text Classification
through
Machine Learning
David W. Miller
Semantic Web
Spring 2002
Department of Computer Science
University of Georgia
www.cs.uga.edu/~miller/SemWeb
Query to General-Purpose Search Engine:
+camp +basketball “north carolina” “two weeks”
Automatic Text Classification through Machine Learning, McCallum, et. al.
2
Domain-Specific
Search Engine
Automatic Text Classification through Machine Learning,
McCallum, et. al.
3
Automatic Text Classification through Machine Learning, McCallum, et. al.
4
Automatic Text Classification through Machine Learning, McCallum, et. al.
5
Domain-Specific Search Engine
Advantages
• High precision.
• Powerful searches on domain-specific
features.
– by location, time, price, institution.
• Domain-specific presentation interfaces:
– Topic hierarchies.
– Specific fields shown in clear format.
– Links for special relationships.
Automatic Text Classification through Machine Learning, McCallum, et. al.
6
Domain-Specific Search Engine
Disadvantages
• Much human effort to build and
maintain!
– e.g. Yahoo has hired many people to build
their hierarchy, and maintain “Full Coverage”,
etc.
Automatic Text Classification through Machine Learning, McCallum, et. al.
7
Tough Tasks
• Find pages that belong in the search engine.
• Find specific fields (price, location, etc).
• Organize the content for browsing.
Automatic Text Classification through Machine Learning, McCallum, et. al.
8
Machine Learning to the Rescue!
• Find pages that belong in the search engine.
– Efficient spidering by reinforcement learning.
• Find specific fields (price, location, etc).
– Information extraction with hidden Markov
models.
• Organize the content for browsing.
– Populate a topic hierarchy by document
classification.
Automatic Text Classification through Machine Learning, McCallum, et. al.
9
Building Text Classifiers
• Manual approach
– Interactive query refinement
– Expert system methodologies
• Supervised learning
1. “Expert” labels example texts with
classes
2. Machine learning algorithm produces
rule that tends to agree with expert
classifications
Machine Learning for Text Classification, David D. Lewis, AT&T Labs
10
Advantages of Using Machine
Learning to Build Classifiers
• Requires no linguistic or computer skills
• Competitive with manual rule-writing
• Forces good practices
– Looking at data
– Estimating accuracy
• Can be combined with manual
engineering
– ML research pays too little attention to this
Machine Learning for Text Classification, David D. Lewis, AT&T Labs
11
Main Processes for a
Machine-Learning System
Prepare
training
samples
Feature
selection
Selected
features
Text
representation
Profiles
/Rules
Model
induction
Feature vectors
Selected features
Supervised Machine-Learning Based Text Categorization, Ng Hong I
12
Preparation of Training Texts
• Essential for a supervised machine
learning text categorization system
• Decide on the set of categories
• A set of positive training texts is
prepared for each of the categories
• Assign subject code(s) to each of the
training texts
– More than one subject code may be
assigned to one training text
Supervised Machine-Learning Based Text Categorization, Ng Hong I
13
Demonstration System: Cora
• Find pages that belong in the search engine.
– Spider CS departments for research papers.
• Find specific fields (price, location, etc).
– Extract titles, authors, abstracts, institutions, etc
from paper headers and references.
• Organize the content for browsing.
– Populate a hand-built topic hierarchy by using text
classification.
Automatic Text Classification through Machine Learning, McCallum, et. al.
14
Automatic Text Classification through Machine Learning, McCallum, et. al.
15
Automatic Text Classification through Machine Learning, McCallum, et. al.
16
See also CiteSeer
[Bollacker, Lawrence
& Giles ‘98]
Automatic Text Classification through Machine Learning, McCallum, et. al.
17
Automatic Text Classification through Machine Learning, McCallum, et. al.
18
Automatic Text Classification
via Statistical Methods
Text Categorization is the problem of assigning
predefined categories to free text documents.
Popular Approach is Statistical Learning Methods
•Bayes Method
•Rocchio Method (most popular)
•Decision Trees
•K-Nearest Neighbor Classification
•Support Vector Machines (fairly new concept)
19
A Probabilistic Generative Model
• Define a probabilistic generative model for
“Bag-of-words”
documents with classes.
Bayes:
Reinforcement
Learning:
a Survey
This paper surveys
the field of reinforcement learning
from a computer
science perspective.
Automatic Text Classification through Machine Learning, McCallum, et. al.
35
1
12
4
1
7
44
3
2
1
5
9
2
56
11
1
…
a
block
computer
field
leg
machine
of
paper
perspective
rate
reinforcement
science
survey
the
this
underrated
…
20
Bayes Method
Pick the most probable class, given the evidence:
c  arg max c j Pr(c j | d )
cj
- a class (like “Planning”)
d
- a document (like “language intelligence proof...”)
Bayes Rule:
Pr(c j | d ) 
Pr(c j ) Pr( d | c j )
Pr( d )
Probability Category cj
should be assigned to
document d
Automatic Text Classification through Machine Learning, McCallum, et. al.
21
Bayes Rule
Pr(c j | d ) 
P(c j | d )
P (d )
Pr(c j ) Pr( d | c j )
Pr( d )
- Probability that document d belongs to category cj
- Probability that a randomly picked document has the same attributes
P(c j )
P(d j | c)
- Probability that a randomly picked document belongs to this category
- Probability that category c contains document d
22
Bayes Method
• Generates conditional probabilities of
particular words occurring in a document
given it belongs to a particular category.
• Larger vocabulary generate better
probabilities
• Each category is given a threshold p for
which it judges the worthiness of a document
to fall in that classification.
• Documents may fall into one, more than one,
or not even one category.
23
Rocchio Method
• Each document is D is represented as a
vector within a given vector space V:

(1)
(| F |)
d  (d ,...,d )
•Documents with similar content have similar
vectors
•Each dimension of the vector space represents a
word selected via a feature selection process
24
Rocchio Method
• Values of d(i) for a document d are
calculated as a combination of the
statistics TF(w,d) and DF(w)
• TF(w,d) (Term Frequency) is the
number of times word w occurs in a
document d.
• DF(w) (Document Frequency) is the
number of documents in which the word
w occurs at least once.
25
Rocchio Method
• The inverse document frequency is
calculated as
IDF(w)  log(
|D|
DF ( w)
)
• Value of d(i) of feature wi for a document d is
calculated as the product
d (i )  TF (wi , d )  IDF(wi )
•d(i) is called the weight of the word wi in the
document d.
26
Rocchio Method
• Based on word weight heuristics, the
word wi is an important indexing term
for a document d if it occurs frequently
in that document
• However, words that occurs frequently
in many document spanning many
categories are rated less importantly
27
Decision Tree Learning
Algorithm
• Probabilistic methods have been
criticized since they are not easily
interpreted by humans, not so with
Decision Trees
• Decision Trees fall into the category of
symbolic (non-numeric) algorithms
28
Decision Trees
• Internal nodes are labeled by terms
• Branches (departing from a node) are
labeled by tests on the weight that the
term has in a test document
• Leafs are labeled by categories
29
Decision Tree Example
30
Decision Tree
• Classifier categorizes a test document d
by recursively testing for the weights
that the terms labeling the internal
nodes have until a leaf node is reached.
• The label of the leaf node is then
assigned to the document
• Most decision trees are binary trees
31
Decision Tree
• Fully grown trees tend to have decision
rules that are overly specific and are
therefore unable to categorize
documents
– Therefore, pruning and growing methods
for such Decision Trees are normally
standard part of the classification packages
32
K-Nearest Neighbor
• Features
– All instances correspond to points in an ndimensional Euclidean space
– Classification is delayed till a new instance
arrives
– Classification done by comparing feature
vectors of the different points
– Target function may be discrete or realvalued
K-Nearest Neighbor Learning, Dipanjan Chakraborty
33
1-Nearest Neighbor
K-Nearest Neighbor Learning, Dipanjan Chakraborty
34
K-Nearest Neighbor
• An arbitrary instance is represented by
(a1(x), a2(x), a3(x),.., an(x))
– ai(x) denotes features
• Euclidean distance between two instances
d(xi, xj)=sqrt (sum for r=1 to n (ar(xi) - ar(xj))2)
• Find the k-nearest neighbors whose distance
from your test cases falls within a threshold p.
• If x of those k-nearest neighbors are in
category ci, then assign the test case to ci,
else it is unmatched.
K-Nearest Neighbor Learning, Dipanjan Chakraborty
35
Support Vector Machines
• Based on the Structural Risk
Minimization principle form
computational learning theory
– Find a hypothesis h for which we can
guarantee the lowest true error
• The true error of h is the probability that h will
make an error on an unseen and randomly
selected test example
36
Evaluating Learning
Algorithms and Software
• How effective/accurate is classification?
• Compatibility with operational
environment
• Resource usage
• Persistence
• Areas learning algorithms need
improvement
Machine Learning for Text Classification, David D. Lewis, AT&T Labs
37
Effectiveness: Contingency
Table
Truth
Yes
No
Yes
a
b
No
c
d
System
Machine Learning for Text Classification, David D. Lewis, AT&T Labs
38
Effectiveness Measures
Truth
Yes
No
Yes
a
b
No
c
d
System
•
•
•
•
•
•
recall = a/(a+c)
precision = a/(a+b)
accuracy = (a+c)/(a+b+c+d)
utility = any weighted average of a,b,c,d
F-measure = 2a/(2a+b+c)
others
Machine Learning for Text Classification, David D. Lewis, AT&T Labs
39
Effectiveness: How to Predict
• Theoretical gaurantees rarely useful
• Test system on manually classified data
– Representativeness of sample important
– Will data vary over time?
– Effectiveness varies widely across classes
and data sets
• Interindexer agreement an upper
bound?
Machine Learning for Text Classification, David D. Lewis, AT&T Labs
40
Effectiveness: How to Improve
• More training data
• Better training data
• Better text representation
– Usual IR tricks (term weighting, etc.)
– Manually construct good predictor features
• e.g. % capitalized letters for spam filtering
• Hand off hard cases to human being
Machine Learning for Text Classification, David D. Lewis, AT&T Labs
41
Conclusions
• Performance of classifier depends
strongly on the choice of data used for
evaluation.
• Dense category space become
problematic for unique categorization,
many documents share characteristics
42
Credits
*This Presentation is Partially Based
on Those of Others Listed Below*
• Supervised Machine Learning Based Text
Categorization
• Machine Learning for Text Classification
• Automatically Building Internet Portals using Machine
Learning
• Web Search
• Machine Learning
• K-Nearest Neighbor Learning
Full Presentations can be found at:
http://webster.cs.uga.edu/~miller/SemWeb/Presentation/ACT.html
43
Resources
• Text Categorization Using Weight Adjusted k-Nearest
Neighbor Classification
• A Probalisitic Analysis of the Rocchio Alg. w/ TFIDF
for Text Categorization
• Text Categorization w/ Support Vector Machines
• Learning to Extract Symbolic Knowledge from the
WWW
• An Evaluation of Statistical Approaches to Text
Categorization
• A Comparison of Two Learning Algorithms for Text
Categorization
• Machine Learning in Automated Text Categorization
Full List of Resources can be found at:
http://webster.cs.uga.edu/~miller/SemWeb/Presentation/ACT.html
44