Classification of Documents

using Text Mining Package “tm”

Pavel Brazdil
LIAAD - INESC Porto LA
FEP, Univ. of Porto

Escola de verão
Aspectos de processamento da LN
F. Letras, UP, 4th June 2009
http://www.liaad.up.pt

Overview

1.Introduction
2.Preprocessing document collection in “tm”
2.1 The dataset 20Newsgroups
2.2 Creating a directory with a corpus for 2 subgroups
2.3 Preprocessing operations
2.4 Creating Document-Term matrix
2.5 Converting the Matrix into Data Frame
2.6 Including class information
3. Classification of documents
3.1 Using a kNN classifier
3.2 Using a Decision Tree classifier
3.3 Using a Neural Net

2
1. Introduction

Package “tm” of R permits to process text documents

in an effective manner
The work with this package can be initiated using a command
> library(tm)
It permits to:
- Create a corpus – a collection of text documents
- Provide various preprocessing operations
- Create a Document-Term matrix
- Inspect / manipulate the Document-Term matrix
(e.g. convert into a data frame needed by classifiers)
- Train a classifier on pre-classified Document-Term data frame
- Apply the trained classifier on new text documents
to obtain class predictions and evaluate performance
3

2 Classification of documents
2.1 The dataset 20Newsgroups

This data is available from

http://people.csail.mit.edu/jrennie/20Newsgroups/
There are two directores:
- 20news-bydate-train (for training a classifier)
- 20news-bydate-test (for applying a classifier / testing)
Each contains 20 directories, each containing the text documents
belonging to one newsgroup
The data can be copied to a PC you are working on
(it is more convenient)

4
20Newsgroups

Subgroup “comp” Subgroup “sci”

comp.graphics sci.crypt
comp.os.ms-windows.misc sci.electronics <- chosen here
comp.sys.ibm.pc.hardware sci.med
comp.sys.mac.hardware sci.space
comp.windows.x Subgroup “talk.politics”
Subgroup “misc” talk.politics.guns
misc.forsale talk.politics.mideast
talk.politics.misc
Subgroup “rec”
rec.autos Subgroup “religion”
rec.motorcycles talk.religion.misc <- chosen here
rec.sport.baseball alt.atheism
rec.sport.hockey soc.religion.christian

2.2 Creating a Corpus

This involves :
- Selecting one of the newsgroups (e.g.sci.electronics)
- Invoking “Change directory” in R to 20news-bydate-train
- Invoking the instruction Corpus():
sci.electr.train <- Corpus( DirSource (“sci.electronics”),
readerControl=list(reader=readNewsgroup, language=“en_US” ) )
If we type :
> sci.electr.train or length(sci.electr.train)
the system responds “A corpus with 591 documents”
Similarly, we can obtain documents from another class
(e.g. talk.religion.misc):
talk.religion.train (377 documents)
Similarly, we can obtain the test data:
sci.electr.test (393 documents)
talk.religion.test (251 documents)
6
Example of one document

sci.electr.train[[1]]
An object of class “NewsgroupDocument”
[1] "In article <[email protected]>
[email protected] writes:"
[2] ">[...]"
[3] ">There are a variety of water-proof housings I could use but the real meat"
[4] ">of the problem is the electronics...hence this posting. What kind of"
[5] ">transmission would be reliable underwater, in murky or even night-time"
[6] ">conditions? I'm not sure if sound is feasible given the distortion under-"
[7] ">water...obviously direction would have to be accurate but range could be"
[8] ">relatively short (I imagine 2 or 3 hundred yards would be more than enough)"
[9] ">"
[10] ">Jim McDonald"
…
[35] " ET \"Tesla was 100 years ahead of his time. Perhaps now his time comes.\""
[36] "----"

Metadata accompanying the document

Slot "Newsgroup": character(0)

Slot "URI": file("sci.electronics/52434", encoding = "UTF-8")
Slot "Cached": [1] TRUE
Slot "Author": [1] "[email protected] (Eric H. Taylor)"
Slot "DateTimeStamp": character(0)
Slot "Description": [1] ""
Slot "ID": [1] "1"
Slot "Origin": character(0)
Slot "Heading": [1] "Re: HELP_WITH_TRACKING_DEVICE"
Slot "Language": [1] "en_US"
Slot "LocalMetaData": list()

See Annex 1 for more details on metadata (slots) etc.

8
2.3 Preprocessing

The Objective of Preprocessing:

Documents are normally represented using words, terms or concepts.
Considering all possible words as potential indicators of a class
can create problems in training a given classifier.
It is desirable to avoid building a classifier using
dependencies based on too few cases (spurious regularities).
The aim of preprocessing is to help to do this.

The function tmMap (available in “tm”) can be used

to carry out various preprocessing steps.
The operation is applied to the whole corpus
(there is not need to program this using a loop).

Preprocessing using tmMap

The format of this function is as follows:

tmMap(Corpus, Function)
The second argument Function determines what is to be done:
asPlain – removes XML from the document,
removeSignature – removes the author of the message,
removeWords, stopwords(language=‘english’)
– removes stopwords for the language specified
stripWhitespace – removes extra spaces,
tmTolower – transforms all upper case letters to lower case,
removePunctuation – removes punctuation symbols,
removeNumbers – removes numbers,
Example of use:
> sci.electr.train <- tmMap(sci.electr.train, asPlain)
> sci.electr.train <- tmMap (sci.electr.train, removeSignature)
etc.
This can be repeated for the other 3 collections of documents 10
Merging document collections

Instead of repeating this for all 4 documents collections

we can merge the four document collections and
perform the preprocessing on the resulting large collection only once.
This can be done using the function c():
> sci.rel.tr.ts <- c(sci.electr.train, talk.religion.train, sci.electr.test, talk.religion.test)
> sci.rel.tr.ts
A text document collection with 1612 text documents.

We need to remember the indices of each document sub-collection

to be able to separate the document collections later.
sci.electr.train – documents 1 .. 591
talk.religion.train – documents 592 .. 968 (377 docs)
sci.electr.test – documents 969 .. 1361 (393 docs)
talk.religion.test – documents 1362 .. 1612 (251 docs)
One single collection is important for the next step (document-term matrix)
11

Preprocessing the entire document collection

> sci.rel.tr.ts.p <- tmMap(sci.rel.tr.ts, asPlain)

> sci.rel.tr.ts.p <- tmMap(sci.rel.tr.ts.p, removeSignature)
> sci.rel.tr.ts.p <- tmMap(sci.rel.tr.ts.p, removeWords,
stopwords(language="english"))
> sci.rel.tr.ts.p <- tmMap(sci.rel.tr.ts.p, stripWhitespace)
> sci.rel.tr.ts.p <- tmMap(sci.rel.tr.ts.p, tmTolower)
> sci.rel.tr.ts.p <- tmMap(sci.rel.tr.ts.p, removePunctuation)
> sci.rel.tr.ts.p <- tmMap(sci.rel.tr.ts.p, removeNumbers)

12
Result of preprocessing of one document

Original document:
> sci.rel.tr.ts[[1]] (=sci.electr.train[[1]])
An object of class “NewsgroupDocument”
[1] "In article <[email protected]>
[email protected] writes:"
[2] ">[...]"
[3] ">There are a variety of water-proof housings I could use but the real meat"
[4] ">of the problem is the electronics...hence this posting. What kind of"
[5] ">transmission would be reliable underwater, in murky or even night-time"
[6] ">conditions? I'm not sure if sound is feasible given the distortion under-"
Pre-processed document:
undesirable
> sci.rel.tr.ts.p[[1]]
[1] in article fbaeffaesoprutgersedu mcdonaldaesoprutgersedu writes
[2]
[3] there variety waterproof housings real meat
[4] of electronicshence posting
[5] transmission reliable underwater murky nighttime
[6] conditions sound feasible distortion under 13

2.4 Creating Document-Term Matrix (DTM)

Existing classifiers that exploit propositional representation,

(such as kNN, NaiveBayes, SVM etc.)
require that data be represented in the form of a table, where:
each row contains one case (here a document),
each column represents a particular atribute / feature (here a word).

The function DocumentTermMatrix(..) can be used to create such a table.

The format of this function is:
DocumentTermMatrix(<DocCollection>, control=list(<Options>))
Simple Example:
> DocumentTermMatrix(sci.rel.tr.ts.p)

Note:
This command is available only in R Version 2.9.1
In R Version 2.8.1 the function available is TermDocMatrix(<DocCollection>)

14
Creating Document-Term Matrix (DTM)

Simple Example:
> DocumentTermMatrix(sci.rel.tr.ts.p)
A document-term matrix (1612 documents, 21967 terms)

Non-/sparse entries: 121.191/35.289.613

Sparsity : 100% problematic

Maximal term length: 135

Weighting : term frequency (tf)

15

Options of DTM

Most important options of DTM:

weighting=TfIdf weighting is Tf-Idf
minWordLength=WL the minimum word length is WL
minDocFreq=ND each word must appear at least in ND docs

Other options of DTM

These are not really needed, if preprocessing has been carried out:
stemming = TRUE stemming is applied
stopwords=TRUE stopwords are eliminated
removeNumbers=True numers are eliminated

Improved example:
> dtm.mx.sci.rel <- DocumentTermMatrix( sci.rel.tr.ts.p,
control=list(weighting=weightTfIdf, minWordLength=2, minDocFreq=5))

16
Generating DTM with different options

> dtm.mx.sci.rel <- DocumentTermMatrix( sci.rel.tr.ts.p,

control=list(minWordLength=2, minDocFreq=5))
> dtm.mx.sci.rel
A document-term matrix (1612 documents, 1169 terms)
Non-/sparse entries: 2.237/1.882.191
Sparsity : 100% better
Maximal term length: 26
Weighting : term frequency (tf)

> dtm.mx.sci.rel.tfidf <- DocumentTermMatrix( sci.rel.tr.ts.p,

control=list(weighing=weightTfIdf, minWordLength=2, minDocFreq=5))
> dtm.mx.sci.rel.tfidf
A document-term matrix (1612 documents, 1169 terms)
Non-/sparse entries: 2.237/1.882.191
Sparsity : 100%
Maximal term length: 26
Weighting : term frequency - inverse document frequency (tf-idf)
17

Inspecting the DTM

Function dim(DTM) permits to obtain the dimensions of the DTM matrix.

Ex.
> dim(dtm.mx.sci.rel)
1612 1169

Inspecting some of the column names:

(ex. 10 columns / words starting with column / word 101)
> colnames(dtm.mx.sci.rel) [101:110]
[1] "blinker" "blood" "blue" "board" "boards" "body" "bonding" "book" "books"
[10] "born"

18
Inspecting the DTM

Inspecting a part of the DTM matrix:

(ex. the first 10 documents and 20 columns)
> inspect( dtm.mx.sci.rel )[1:10,101:110]
Docs blinker blood blue board boards body bonding book books born
1 0 0 0 0 0 0 0 0 0 0
2 0 0 0 0 0 0 0 0 0 0
3 0 0 0 0 0 0 0 0 0 0
4 0 0 0 0 0 0 0 0 0 0
5 0 0 0 0 0 0 0 0 0 0
6 0 0 0 0 0 0 0 0 0 0
7 0 0 0 0 0 0 0 0 0 0
8 0 0 0 0 0 0 0 0 0 0
9 0 0 0 0 0 0 0 0 0 0
10 0 0 0 0 0 0 0 0 0 0

As we can see, the matrix is very sparse. By chance there are no values other than 0s.
Note: The DTM is not and ordinary matrix, as it exploits
object-oriented representation (includes meta-data).
The function inspect(..) converts this into an ordinary matrix which can be inspected.19

Finding Frequent Terms

The function findFreqTerms(DTM, ND) permits to find

all the terms that appear in at least ND documents.

Ex.
> freqterms100 <- findFreqTerms( dtm.mx.sci.rel, 100)
> freqterms100
[1] "wire" "elohim" "god" "jehovah" "lord" From talk.religion

> freqterms40 <- findFreqTerms(dtm.mx.sci.rel, 40)

> freqterms40
[1] "cable" "circuit" "ground" "neutral" "outlets" "subject" "wire" "wiring"
[9] "judas" "ra" "christ" "elohim" "father" "god" "gods" "jehovah"
[17] "jesus" "lord" "mcconkie" "ps" "son" "unto"

> freqterms10 <- findFreqTerms(dtm.mx.sci.rel, 10)

> length(freqterms10)
[1] 311
20
2.5 Converting TDM into a Data Frame

Existing classifiers in R (such as kNN, NaiveBayes, SVM etc.)

require that data be represented in the form of a data frame
(particular representation of tables).
So, we need to convert DT matrix into a DT data frame:
> dtm.sci.rel <- as.data.frame(inspect( dtm.mx.sci.rel ))
> rownames(dtm.sci.rel)<- 1:nrow(dtm.mx.sci.rel)
sci.electr portion
> dtm.sci.rel$wire[180:200]
[1] 0 6 0 8 108 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
> dtm.sci.rel$god[180:200] talk.religion portion
[1] 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0

> dtm.sci.rel$wire[(592+141):(592+160)] sci.electr portion

[1] 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
> dtm.sci.rel$god[(592+141):(592+160)] talk.religion portion
[1] 0 0 0 10 0 5 0 0 9 0 0 35 0 0 0 0 0 0 0 0

21

Converting TDM into a Data Frame

We repeat this also for the tf.idf version:

So, we need to convert DT matrix into a DT data frame:

> dtm.sci.rel.tfidf <- as.data.frame(inspect(dtm.mx.sci.rel.tfidf))

> round(dtm.sci.rel.tfidf$wire[180:200] ,1) sci.electr portion

[1] 0.0 8.3 0.0 8.3 8.3 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
> round(dtm.sci.rel.tfidf$god[180:200],1) talk.religion portion
[1] 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0

> round(dtm.sci.rel.tfidf $wire[(592+141):(592+160)] ,1) sci.electr portion

[1] 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
> round(dtm.sci.rel.tfidf$god[(592+141):(592+160)],1) talk.religion portion
0.0 0.0 0.0 5.1 0.0 5.1 0.0 0.0 5.1 0.0 0.0 5.1 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0

22
2.6 Appending class information

This includes two steps:

- Generate a vector with class information,
- Append the vector as the last column to the data frame.

Step 1. Generate a vector with class values (e.g. “sci”, “rel”)

We know that (see slide 6):
sci.electr.train – 591 docs talk.religion.train – 377 docs
sci.electr.test – 393 docs talk.religion.test - 251 docs
So:
> class <- c(rep(“sci”,591), rep(“rel”,377), rep(“sci”,393), rep(“rel”,251))
> class.tr <- c(rep("sci",591), rep("rel",377))
> class.ts <- c(rep("sci",393), rep("rel",251))
Step2. Append the class vector as the last column to the data frame
> dtm.sci.rel <- cbind( dtm.sci.rel, class)
> ncol(dtm.sci.rel)
[1] 1170 (the number of columns has increased by 1)
23

3 Classification of Documents
3.1 Using a KNN classifier

Preparing the Data

The classifier kNN in R requires that
the training data and test data have the same size.
So:
sci.electr.train – 591 docs : Use the first 393 docs
talk.religion.train – 377 docs: Use the first 251 docs
sci.electr.test – 393 docs: Use all
talk.religion.test - 251 docs: Use all
> l1 <- length(sci.electron.train) (591 docs)
> l2 <- length(talk.religion.train) (377 docs)
> l3 <- length(sci.electron.test) (393 docs)
> l4 <- length(talk.religion.test) (251 docs)

> m1 <- min(l1,l3) (393 docs)

> m2 <- min(l2,l4) (251 docs)
> m3 <- min(l1,l3) (393 docs)
> m4 <- min(l2,l4) (251 docs)
24
Preparing the training data

Generating the training data:

Calculate the last column of data (excluding the class):
> last.col <- ncol( dtm.sci.rel )-1
Generate the first part of the training data with lines of sci.electr.train
> sci.rel.tr <- dtm.sci.rel[1: m1, 1:last.col]
Generate a vector with the class values:
> class.tr <- dtm.sci.rel[1: m1, last.col+1]
Extend the training data by adding lines from talk.religion.train
> sci.rel.k.tr[(m1+1):(m1+m2),] <- dtm.sci.rel[(l1+1):(l1+m2),1:last.col]
> nrow(sci.rel.k.tr)
644
Add class values at the end of the class vector:
> class.k.tr[(m1+1):(m1+m2)] <- dtm.sci.rel[(l1+1):(l1+m2), last.col+1]
> length(class.tr)
644

25

Preparing the test data

Generate the test data using

the appropriate lines of sci.electr.test and talk.religion.test
> sci.rel.k.ts <- dtm.sci.rel[(l1+l2+1):(l1+l2+m3+m4),1:last.col]
> nrow(sci.rel.k.ts)
644
> class.k.ts <- dtm.sci.rel[(l1+l2+1):(l1+l2+m3+m4),last.col+1]

26
3.2 Classification of Docs using a Decision Tree

Here we will use the same training / test data as in

> library(rpart)

Here we will use just 20 frequent terms as attributes

> freqterms40
[1] "cable" "circuit" "ground" "neutral" "outlets" "subject" "wire"
[8] "wiring" "judas" "ra" "christ" "elohim" "father" "god"
[15] "gods" "jehovah" "jesus" "lord" "mcconkie" "ps" "son"
[22] "unto"
> dt <- rpart(class ~ cable + circuit + ground + neutral + outlets + subject + wire +
wiring + judas + ra + christ + elohim + father + god + gods + jehovah + jesus +
lord + mcconkie + ps + son + unto, sci.rel.tr)

27

Inspecting the Decision Tree

> dt
n= 968
node), split, n, loss, yval, (yprob)
* denotes terminal node
1) root 968 377 sci (0.3894628 0.6105372)
2) god>=2.5 26 0 rel (1.0000000 0.0000000) *
3) god< 2.5 942 351 sci (0.3726115 0.6273885)
6) jesus>=2.5 16 0 rel (1.0000000 0.0000000) *
7) jesus< 2.5 926 335 sci (0.3617711 0.6382289) *

28
Evaluating the Decision Tree

Decision Tree Test data

> dt.predictions.ts <- predict(dt, sci.rel.ts, type="class")

> table(class.ts, dt.predictions.ts)
dt.predictions.ts
class.ts rel sci
rel 22 229
sci 0 393

29

3.3 Classification of Documents using a Neural Net

Acknowledgements:
Rui Pedrosa, M.Sc. Student, M. ADSAD, 2009

> nnet.classifier <- nnet(class ~., data=sci.rel.tr, size=2, rang=0.1,

decay=5e-4, maxit=200)

> predictions <- predict(nnet.classifier, sci.rel.ts, type=“class”)

The errors reported on a similar task were quite good - about 17%

30
5. Calculating the evaluation measures

The basis for all calculations is the confusion matrix, such as:
> conf.mx <- table(class.ts, predictions.ts)

> conf.mx
class.ts rel sci
rel 201 50
sci 3 390

> error.rate <- (sum(conf.mx) – diag(conf.mx)) / sum(conf.mx)

> tp <- conf.mx[1,1] (true positives)

> fp <- conf.mx[2,1] (false positives)
> tn <- conf.mx[2,2] (true negatives)
> fn <- conf.mx[1,2] (false negatives)

> error.rate <- (tp + tn) / (tp + tn + fp + fn)

31

Evaluation measures

+^ -^
+ TP FN
Recall = TP / (TP + FN) * 100%
- FP TN

Precision = TP / (TP + FP) * 100% +^ -^

+ TP FN

- FP TN

> recall = tp / (tp + fn)

> precision = tp / (tp + fp)
> f1 = 2 * precision * recall / (precision + recall)

32