Text Analytics Teradata & Sabanci University April, 2015.

Text Analytics

Teradata & Sabanci University

April, 2015

2

TEXTDATA

#2 Sentiment

Analysis

#3Relation between

words

#4 Association Rules between events

#1 Topic Categorization

Text analytics scope at a glance

3

Main Project Diagram

4

#1 Topic Categorization

5

2 Main Categories

Şikayet

Bilgi/İşlem

19 Subcategorie

s

Ürün/Servis

Kampanya/ Paket

.........

Fatura ve Yükleme

Topic Categorization

6

Dimension Reduction◦ Data

◦ High in volume

◦ Noisy

◦ For each category, significant words found (by using Aster tf_idf function)

◦ Project the data into the space of these words:

◦ Decrease the noise

◦ Thus, differentiate the classes better

SAMPLE

Topic Categorization

7

Main Category Classification

We applied Naive Bayes Text Classification

CE660n660l

CE660n660l

CE660n660l

2-class Classification

Correct PredictionIncorrect Prediction

8

Subcategory Classification

We applied Naive Bayes Text Classification

CE660n660l

CE660n660l

CE660n660l

19-class Classification

Correct PredictionIncorrect Prediction

9

2 Main Categories 19 Subcategories

Topic Categorization

10

#2 Sentiment Analysis

11

Sentiment Classification Model

• Whole dataset is randomly splitted as train and test sets• The model is unbiasedly evaluated on the unseen test set• Aster’s Random Forest Function is available to apply to our Turkish Sentiment Features.

12

Turkish Model

13

Does this call have NEGATIVE sentiment?

• Classify the call center messages into two:– Negative

– Neutral (and Positive)

• Train and test sets are different for each evaluation:– The accuracy is not dependent on a specific test set

Classification

correct misclassified

> 81

<19

14

Regression Model

• Same dataset used for the sentiment classification

• Output is different; not classes, but real values

• These values scale the negativity of the customers

15

Negative Sentiment Breakdown

2%

93%

5%

highmediumlow

• Threshold values were chosen by using 68–95–99.7 rule in statistics (three-sigma rule of thumb heuristic supports this).

• Our thresholds are set as mean+(-)2*standart deviation

• Used for social sciences and business

16

#3 Relation Extraction

17

Relations between Words

• Relation scores are calculated between words

• It is based on confidence values.– Score(a,b) = confidence(a,b)*confidence(b*a)

Words were filtered:– We selected the words with the highest tf*idf values

Words are clustered in terms of their relations

Clusters are represented in a graph Cfilter (in Aster)function has been applied for this process.

18

Complaining Customers Graph

Selected Cluster

19

Closer Look to the Cluster

The cluster seems to indicate the customers facing internet-related problems

• Connection problem, or• Internet package is not sufficient?

20

Complaint Word Cloud (Aster Lens)

21

Information&Operation Word Cloud

22

Bigram Cloud (ngram function)

23

#4 Association Rules between Events

24

Association Rule Mining

Topic A with a Sentiment score

Topic B with a Sentiment score

Topic C

Confidence(A, B --> C)

25

QUESTIONS