1. A Straightforward Author Profiling Approach in MapReduce
Suraj Maharjan, Prasha Shrestha, Thamar Solorio and Ragib Hasan

2. ? Introduction Author profiling
identifying age-group, gender, native language, personality and other aspects that constitute the profile of an author, by analyzing his/her writings ?

3. Applications Forensics and Security
Threatening
s containing spam/malwares
Authorship attribution of old texts
Not possible to check against every author
Profiling narrows down the list
Marketing
Find out demographics/profiles of customers who either love or hate their product.
Find out target group of customers.

4. Dataset Age Gender English Spanish Training Early Bird Test 10s male
8600
740
888
1250
120
144
female
20s
42900
3840
4608
21300
1920
2304
30s
66800
6020
7224
15400
1360
1632
Total
236600
21200
25440
75900
6800
8160
`Balanced across gender; imbalanced across age
Table 1: PAN’13 Training, Early bird, Test documents distribution.

5. Dataset English 1.8 GB 135 MB 168 MB Spanish 384 MB 34 MB 40 MB Total
Language
Training
Early Bird
Test
English
1.8 GB
135 MB
168 MB
Spanish
384 MB
34 MB
40 MB
Total
2.2 GB
169 MB
209 MB
Table 2: PAN’13 Data Size

6. Why MapReduce?
Provides abstraction for designing distributed algorithms
compared to MPI, p-threads
No need to worry about
deadlocks,
race conditions
machine failures

7. Methodology Preprocessing Features Classification Algorithm
Sequence File Creation
Tokenization
DF Calculation
Filter
Features
Word n-grams (unigrams, bigrams, trigrams)
Weighing Scheme: TF-IDF
Classification Algorithm
Naïve Bayes

8. Tokenization Job Input Map Output
Key: filename (<authorid>_<lang>_<age>_<gender>.xml)
Value: content
Map
Lowercase
1, 2, 3-grams (Lucene)
Output
Key: filename
Value: list of ngram tokens

9. Tokenization Remove xml and html tags from documents
<authorid>_<lang>_<age>_<gender>.xml
Preprocessing
Filename
Content
Filename
1,2,3 grams
Sequence Files F1
“A B C C” F2
“B D E A” F3
“A B D E” F4
“C C C” F5
“B F G H” F6
“A E O U” F1
[A, B, C, C, A B, A B C, ..] F2
[B, D, E, A, B D E, E A, ..] F3
[A, B, D, E, B D, B D E, …] F4
[C, C, C, C C, C C C, …] F5
[B, F, G, H, G H, …] F6
[A, E, O, U, E O, A U,…]
Map
Tokenization job

10. DF Calculation Job Mapper Reducer Combiner
Map(“<authorid>_<lang>_<age>_<gender>.xml”,[a ab abc ..]) ->list(token,1)
Similar to word count but need to emit only unique token, once per document
Reducer
Reduce(token, list(1,1,1..)) –>(token, df_count)
Combiner
Minimizes network data transfer

11. DF Calculation Job Reduce Token 1 Token 1 Token DF count DF count job… 1 B D C A 1 … B F G .. A 4 B C 2 D E 3 F 1
DF count job
Group By

12. Filter Job Filters out the least frequent tokens base on DF scores
Build dictionary file
Maps each token with unique integer id
Mapper
Setup
Read in the dictionary file, df scores
Map
map(“<authorid>_<lang>_<age>_<gender>.xml”, token list) -> (“<authorid>_<lang>_<age>_<gender>.xml”, filtered token list)

13. Filter Job Filter count DF count Filename 1,2,3 grams Map Filter job
[A, B, C, C, A B, A B C, ..] F2
[B, D, E, A, B D E, E A, ..] F3
[A, B, D, E, B D, B D E, …] F4
[C, C, C, C C, C C C, …] F5
[B, F, G, H, G H, …] F6
[A, E, O, U, E O, A U,…]
Map F1
[A, B, C, C, A B, ..] F2
[B, D, E, A, B D E, ..] F3
[A, B, D, E, B D, B D E, …] F4
[C, C, C, C C, …] F5
[B, H, …] F6
[A, E,…]
Filter job

14. TF-IDF Job Mapper Setup: Map: Read in dictionary and DF score files
Map(“<authorid>_<lang>_<age>_<gender>.xml”, filtered token list)- >(“<authorid>_<lang>_<age>_<gender>.xml”, VectorWritable)
Compute tf-idf scores for each token
Creates RandomSparseVector(mahout-math)
Finally writes vectors

15. Training Naïve Bayes (MR)
Translate TF-IDF vectors into LibLinear feature file
Use LibLinear to train Logistic Regression
Liblinear is also very fast

16. Naïve Bayes MR Mapper Reducer Setup Map CleanUp
Initialize a prior probability vector
Map
Compute partial sums and store it in prior probability vector
Map(“authorid_age_gender.xml”,Vector)->(id for age_gender_class,Vector)
CleanUp
Emit(-1,prior probability vector)
Reducer
Vector sum reducer (mahout)

17. Experiments Local Hadoop cluster with 1 master node and 7 slave nodes
Each node has 16 cores and 12 GB memory
Modeled as 6 class classification problem

18. Results Number of features in English : 2,908,894
Number of features in Spanish: 2,806,922
Language
System
Total (%)
Age (%)
Gender (%)
English
Baseline
28.4
56.8 50
PAN 2013 English Best
38.94
59.21
64.91
PAN 2013 Overall Best
38.13
56.9
65.72
Ours (Test)
42.57
65.62
61.66
Ours (Early Bird)
43.88
65.8
62.57
Spanish
28.23
56.47
PAN 2013 Spanish Best
42.08
64.73
64.3
41.58
62.99
65.58
40.32
61.73
64.63
40.62
62.1
64.79
Table 3: Accuracy

19. Results Accuracy(%) Filter at DF # of features 10 2,908,894 42.57 15
1,885,410
42.37 20
1,403,139
42.26
100
276,902
41.93
200
136,999
37.34
300
90,836
17.23
400
67,938
7.8
500
54,330
5.93
1000
27,151
4.66
Table 4: Accuracy on test dataset compared with \# of features for English

20. Results Parameter Setting Accuracy (%) N-gram
Character Bigrams and Trigrams
31.2
%word uni, bi and tri grams
TFIDF
42.57
Word Unigrams
39.99
Word Bigrams
41.17
Word Trigrams
39.7
Stopword Unigrams
29.14
Weighing Scheme TF
36.12
Preprocessing
Stopwords filtered and Porter Stemming
35.82
Table 5: Accuracy comparison for different parameter settings for English dataset

21. Results
Time (Minutes)
Figure 1: Runtimes for English

22. Results
Time (Minutes)
Figure 2: Runtimes for Spanish

23. Conclusion and Future Works
Word n-grams proved to be better features than character n-grams for this task
Stop words proved to be good features
TF-IDF is better than TF
Simple approaches do work
Filter out spam like documents
Add more sophisticated features and make them work in MapReduce

24. Thank you.