Transcript Slides12 (CRM)

CRM Segmentation
Segmentation of Textual Data
Zhangxi Lin
Overview




Text Mining Review
Converting Unstructured Text to Structured Data
Segmenting Textual Data
Demonstrations
2
Text Mining Review
3
Text Mining – Why and How
• The volume of text data is much greater than that of
numeric data
• The means dealing with text data is far from enough
What Text Mining Is
• Text mining is a process that employs a set of
algorithms for converting unstructured text into
structured data objects and the quantitative methods
used to analyze these data objects.
• “SAS defines text mining as the process of
investigating a large collection of free-form documents
in order to discover and use the knowledge that exists
in the collection as a whole.” (SAS Text Miner:
Distilling Textual Data for Competitive Business
Advantage)
5
What Text Mining Is Not
Text mining is not
 a text summarization tool
 an information extraction methodology
 a natural language processor.
6
Two Types of Document Data
Separate
Document Files
(TMFILTER)
Document Text Field
7
The SAS Text Mining Process
1. Preprocess document files to create a SAS data set.
 TMFILTER macro
 SAS language features
2. Parse the document field.
 PARSE tab in Text Miner
 Stemming
 Part-of-speech tagging
 Entities
 Stop/start lists
 Synonym lists
 And so forth
continued...
8
The SAS Text Mining Process
3. Derive the term by document frequency matrix.
 The Text Miner Transform tab
 Frequency weights
 Term weights
4. Transform the term by document frequency matrix.
 The Text Miner Transform Tab
 Singular Value Decomposition (SVD)
 Roll Up Terms
5. Perform the analysis.
 Exploration
 Clustering/unsupervised learning
 Predictive modeling
9
Text Mining Strengths





Clustering documents in a corpus
Investigating word (token) distribution across
documents within a corpus
Identifying words with the highest discriminatory
power
Classifying documents into predefined
categories
Integrating text data with structured data to
enrich predictive modeling endeavors
10
Text Mining Deficiencies
• Text mining algorithms perform poorly in distinguishing
negations, for example:
 Herman was involved in a motor vehicle accident.
 Herman was NOT involved in a motor vehicle accident.
• Text mining cannot generally make value judgments, for
example, classifying an article as positive or negative with
respect to any tokens it contains.
11
Text Mining Deficiencies
• Text mining algorithms do not work well with large
documents.
 Performance is slow.
 Increased term occurrence across documents
decreases separation of documents.
12
The SAS Text Mining Process
1. Preprocess document files to create a SAS data set.
 TMFILTER macro
 SAS language features
2. Parse the document field.
 PARSE tab in Text Miner
 Stemming
 Part-of-speech tagging
 Entities
 Stop/start lists
 Synonym lists
 And so forth
continued...
13
The SAS Text Mining Process
3. Derive the term by document frequency matrix.
 The Text Miner Transform tab
 Frequency weights
 Term weights
4. Transform the term by document frequency matrix.
 The Text Miner Transform Tab
 Singular Value Decomposition (SVD)
 Roll Up Terms
5. Perform the analysis.
 Exploration
 Clustering/unsupervised learning
 Predictive modeling
14
Using the TMFILTER
Macro with the
Newsgroups Data
• This demonstration illustrates how to use the
TMFILTER macro to process groups of text
files.
15
SAS Text Miner Text Processing Features





16
Text parsing
Removal of stop words
Part-of-speech tagging
Stems and synonym handling
Entities
Stop Words
• Stop words are words that have little or no value in
identifying a document or in comparing documents.
Standard stop lists contain stop words that are
 Articles (the, a, this)
 Conjunctions (and, but, or)
 Prepositions (of, from, by).
• Custom stop lists identify low information words, like
the word “computer” in a collection of articles about
computers.
17
Sashelp.stoplst
18
Default Stop Lists
• A default stop list or a user-defined stop list defines
stop words to be removed.
• Default stop lists:
 English: sashelp.stoplst
 French: sashelp.frchstop
 German: sashelp.grmnstop
 Mixed: sashelp.mixdstop
19
Stop List versus Start List


20
Use a start list when
– documents are dominated by “technical jargon”
– domain expertise can enhance text mining.
Use a stop list when
– documents are loosely related: news, business
reports, Internet searches
– domain expertise is not available.
Issues in Creating a Start List


21
Do not just add high frequency terms.
– Low frequency terms that only appear in a few
documents may be good discriminators.
– High frequency terms may be candidates for a stop
list.
Data-derived start lists should be reviewed by domain
experts.
Tagging Parts of Speech
• Determines if the word is a common noun, verb,
adjective, proper noun, adverb, and so forth.
• Disambiguate parts of speech when a word is used in a
different context,
 I wish that my bank had more ATM machines.
 You can bank on either Philadelphia or Oakland
winning the Super Bowl next year.
 Settlers living on the west bank of the river were
forced to relocate.
22
Tagging Parts of Speech in Text Miner
continued...
23
Tagging Parts of Speech in Text Miner
24
Stemming




May employ algorithm and/or table look up
– Porter stemmer
– Levin stemmer
Errors of commission (organizationorgan)
Errors of omission (matricesmatrix)
Can be related to spell checking
continued...
25
Stemming Examples
• BIG: BIG, BIGGER, BIGGEST
• REACH: REACH, REACHES, REACHED, REACHING
• WORK: WORK, WORKS, WORKED, WORKING
• CHILD: CHILD, CHILDREN
• KNIFE: KNIFE, KNIVES
• PERRO: PERRO, PERRA (Spanish, male and female
dog)
26
Stemming in Text Miner
continued...
27
Stemming in Text Miner



28
Text Miner performs stemming to derive stem
synonyms, for example, run/ran/runs/running, and
combines these with defined synonyms, for example,
run/sprint.
The default synonym data set for Text Miner,
sashelp.engsynms, is primarily for illustration.
Synonyms may split based on part of speech, for
example, teach/train=verb, locomotive/train=noun.
Synonyms




Language dictionaries
Technical jargon
Abbreviations
Specialty dictionaries
• Note: This could be associated with stemming in file
preprocessing.
continued...
29
Synonyms
train
instruct
teach
educate
teach
30
Synonym Lists


Default: sashelp.engsynms (ToolsSettings)
User Defined
– SAS Data Set
– Three fields: TERM ($25.), PARENT ($25.),
CATEGORY ($12.)
– Example: TERM=EM, PARENT=Enterprise Miner,
CATEGORY=PRODUCT
continued...
31
Converting Unstructured Text to
Structured Data
32
Term-Document Frequency Matrices
Documents
D1
D2
…
Term
ID
Dn
T1
1
D1,1 D1,2 … D1,n
T2
2
D2,1 D2,2 … D2,n
…
…
Di,j=count of number of times word i occurs in document j
33
Term-Document Frequency Matrices
• Pitfalls
 Sparse cells (many zeroes)
 Weak discriminatory power
 Too large
• Solution
 Term frequency functions
 Singular value decomposition
34
Weighted Term-Document Frequency Matrix
D1
D2
…
Word phrase
ID
Dn
T1
1
W1,1 W1,2 … W1,n
T2
2
W2,1 W2,2 … W2,n
…
…
Wi,j=weight associated with word i in document j
35
Frequency and Term Weights Notation
ai , j  frequency that term i appears in document j
gi  frequency that term i appears in document
collection
n  number of documents in the collection
di  number of documents in which t erm i appears
ai , j
pi , j 
gi
36
Deriving the Weighted Term-Document
Frequency Matrix
aˆij  Gi Lij
aˆij  expected frequency
Gi  term weight
Lij  frequency weight
37
Transformed Term-Document Frequency
Matrix Elements
• The original frequencies
ai , j
• in the Term-Document Frequency Matrix are
transformed to the “expected” frequencies
aˆi , j  Li , j Gi
38
Default Weights
Term Weight=Entropy
Gi  1   j
pij log 2 ( pij )
log 2 (n)
, pij 
aij
gi
, g i   j aij
Frequency Weight=Log
Lij  log 2 (aij  1)
39
Frequency Weights
Binary
1 if term i is in document j
Lij  
0 otherwise
Log
Lij  log 2 (aij  1)
None
Lij  aij
40
Singular Value Decomposition
•
•
•
•
Classical SVD in statistics: A=UV
For term-document frequency matrix A, U is the matrix of term
vectors,  is a diagonal matrix with singular values along the
diagonal, and V is the matrix of document vectors.
The projection V* is output as a set of SVD dimensions for
each document, with the dimensions stored in the variables
COL1, COL2, and so forth. V* is a sub-matrix of V determined
by the maximum dimension specified by the user.
Resolution (low/medium/high) changes the cutoff value for
selecting a “significant” number of dimensions.
41
SVD is very useful for
•
•
•
•
Compression
Noise reduction
Finding ”concepts” or ”topics” (text mining/LSI)
Data exploration and visualizing data (e.g. spatial
data/PCA)
• Classification (of e.g. handwritten digits)
42
SVD appears under different names
• Principal Component Analysis (PCA)
• Latent Semantic Indexing (LSI)/Latent Semantic
Analysis (LSA)
• Karhunen-Loeve expansion/Hotelling transform (in
image processing)
43
Segmenting Textual Data
The Text Mining Project

Document analysis is the goal of the project
– Exploratory analysis of document collections
– Clustering of documents as an aid to human
evaluation of documents
45
Text Mining as Part of a Data Mining Project



Predictive modeling with many fields, one or more
of which are unstructured text
Recommender systems
Others
46
Precision vs. Recall
• Measure to describe how effective a binary text classifier predicts
documents that are relevant to a particular category.
• Precision – The percentage of the predicted positive in all positive
instances
– Precision = TP / (TP + FN)
• Recall – How well the classifier can find relevant documents and
properly assign them to their correct category
– Recall = TP / (TP + FP)
47
Text Mining as Part of a Data Mining Project


The goals of the project influence how text mining is
performed.
A single unstructured text field becomes a set of K
quantitative inputs.
48
Memory-Based Reasoning
• Memory-based reasoning is a process that identifies similar cases
and applies the information that is obtained from these cases to a
new record. In Enterprise Miner, the Memory-Based Reasoning
node is a modeling tool that uses a k-nearest neighbor algorithm to
categorize or predict observations.
• The k-nearest neighbor algorithm takes a data set and a probe,
where each observation in the data set is composed of a set of
variables and the probe has one value for each variable. The
distance between an observation and the probe is calculated. The
k observations that have the smallest distances to the probe are
the k-nearest neighbor to that probe.
49