approaches for source retrieval and text ... - pan.webis.de · pan@clef2013 heilongjiang institute...

Kong Leilei, Qi Haoliang, Du Cuixia, Wang Mingxing, Han Zhongyuan

www.hljit.edu.cnPAN@CLEF2013

Approaches for Source Retrieval and Text Alignment of Plagiarism Detection

1

Who are we?

2

Who are we?

Heilongjiang Institute of TechnologyHarbin, Heilongjiang Province,

China

2

Our University

Heilongjiang Institute of Technology, Kong LeileiPAN@CLEF20133

Index

Approaches for Source Retrieval

Approaches for Text Alignment

Further works


13

Source Retrieval


Candidate Documents

Suspicious plagiarism textDocument

Set

Internet Resource

Text Alignment

SourceRetrieval

Suspicious document

Query keywords

14

Source Retrieval


Candidate Documents


Set

Internet Resource

Text Alignment

SourceRetrieval

Suspicious document

Query keywords

2 problmes of source retrieval

Tow core problem of source retrieval Retrieval source is millions of documents

from the Internet This work was done by PAN

The query keywords of suspicious document which would be used for retrieval are not specified How to extract query keyword is one of important

issues of our work

6

16

Query Keywords Extraction

Query Keywords Extraction Based on TF-IDF

Query Keywords Extraction Based on Weighted TF-IDF

Adjacent Query Keywords Extraction by PatTree

Combination of Queries and Execution of Retrieval


17

Keywords Based on TF-IDF

TF - term frequency, denotes the frequency of term i in document j

IDF - inverse document frequency

IDF =log2 (N/ df j)

TF-IDF of term i is:

Tips: we found that the top 10 terms with the highest TF-IDF can obtain a good results


18

Keywords Based on Weighted TF-IDF

Weighted TF-IDF

Where weight is a weighted parameter, and we calculate the weight of term i according to its location

Tips: the keywords extraction based on the weighted TF-IDF sometimes is useful, sometimes useless.


19

Adjacent Query Keywords Extraction by PatTree

The adjacent string with high frequency is more important than a single word

We use PatTree - an efficient data structure– to get the adjacent strings and their frequency


example

20

Combination of Queries

Query Query Keywords

1

2

3

4

5

6

7

8

9

Top 1 to 5 query keywords based on TF-IDF

Top 2 to 10 query keywords based on TF-IDF

2-Gram query keywords based on PatTree




Top 1 to 5 query keywords based on weighted TF-IDF

Top 6 to 10 query keywords based on weighted TF-IDF



Table 1: Query Combination and Group

21

Results Source Retrieval subtask

Workload

Queries 48.50

Downloads 5691.47

Time to 1st Detection

Queries 2.46

Downloads 285.66

Retrieved Performance

Precision 0.01

Recall 0.65

No Detection 3


Table2：Results Source of PAN@CLEF2013 Retrieval subtask

22


Workload

Queries 48.50

Downloads 5691.47


Queries 2.46

Downloads 285.66


Precision 0.01

Recall 0.65

No Detection 3



23


Workload

Queries 48.50

Downloads 5691.47


Queries 2.46

Downloads 285.66


Precision 0.01

Recall 0.65

No Detection 3



24

Candidate Documents


Set

Internet Resource

Text Alignment

Text Alignment

Source Retrieval


Suspicious document

Query keywords

25

Candidate Documents


Set

Internet Resource

Text Alignment

Text Alignment

Source Retrieval


Suspicious document

Query keywords

26

Text Alignment


27

Text Alignment


28

Text Alignment

Seeding


29

Text Alignment

Seeding


30

Match Merging

Text Alignment

Seeding


31

Match Merging

Text Alignment

Seeding


32

Match Merging

Text Alignment

Extraction FilteringSeeding


33

Match Merging

Text Alignment

Extraction FilteringSeeding


34

Seeding Match Merging

Text Alignment

Extraction Filtering


35


Text Alignment



36


Text Alignment



37


Text Alignment



38


Text Alignment



39


Text Alignment



•Bilateral Alternating

Merging Algorithm

40


Text Alignment



41


Text Alignment



42


Text Alignment



43

Performance on the PAN2012 test corpus

Table 3: Overall evaluation results for the final test corpus


44


Table 4: Results for the 02-no-obfuscation sub-corpus


45


Table 5: Results for the 03-random-obfuscation


46


Table 6: Results for the 04-translation-obfuscation


47


Table 7: Evaluation results for the 05-summary-obfuscation


48

Further work

Use different methods to deal with different plagiarism problems to obtain a better performance

Query keywords extraction and ranking

Thank you for your attention!