Detecting Cross-Language Plagiarism using Open Knowledge Graphs

Corresponding authors: Norman Meuschke, Terry Ruas
Venue: 2nd Workshop on Extraction and Evaluation of Knowledge Entities from Scientific Documents (EEKE2021)
at the ACM/IEEE Joint Conference on Digital Libraries 2021 (JCDL2021)

==========================================================================

Source code: https://github.com/ag-gipp/cl-osa 

==========================================================================

Dataset Details

ASPEC. The Asian Scientific Paper Excerpt Corpus comprises excepts of scientific papers in Japanese that have been manually translated to English and Chinese. We use both subsets of the ASPEC corpus. 

• ASPEC-JC contains abstracts and paragraphs from the main text of research papers that were translated manually from Japanese to Chinese.
• ASPEC-JE contains abstracts of approx. two million research papers that were translated manually from Japanese to English.  

JRC-Acquis. The corpus consists of legislative texts in 22 languages, which the European Union's Joint Research Centre (JRC) selected from the cumulative body of EU laws (the so called Acquis communautaire). We sampled our test cases from the 10,000 document pairs in the English-French subset of the corpus.

Europarl. The corpus contains transcripts of European Parliament proceedings in 21 European languages. We exclusively sampled test cases from the 9,443 document pairs in the English-French subset of the corpus.

PAN-PC-11. The corpus contains instances of simulated monolingual and cross-language plagiarism that were used for evaluating plagiarism detection methods as part of the workshop series Plagiarism Analysis, Authorship Identification, and Near-Duplicate Detection (PAN). Most of the 26,939 documents in the corpus were created by extracting text from openly available books. The documents are partially interspersed with instances of simulated plagiarism that were created and obfuscated automatically or by crowdsourced workers. We exclusively sampled test cases from the 2,921 Spanish-English aligned document pairs in the corpus, for which simulated plagiarism instances were either machine-generated or created manually by crowdsourced workers.

==========================================================================

File Structure

[corpus_documents] folder: Corpora of translation-aligned documents used in our experiments composed of: 

• aspec: Japanese and English
• aspecx: Japanese and Chinese 
• jrc: English and French                  
• europarl: English and French
• pan: English and Spanish

Each sub-corpus consists of 4,000 translation-aligned files (2,000 per language); the entire corpus has thus 20,000 files.
Each set of translation-aligned documents was randomly selected from the original datasets (details in the paper).
The Japanese files in aspec and aspecx do not necessarily overlap even though they are from the same dataset.

[vectors_documents] folder: Average vector representation of the documents in the datasets from two pre-trained models:

• Universal Sentence Encoder - Multilingual (USE-ML)
• ConceptNet Numberbatch

Naming convention: <model>_<dataset>_<language>;

• Example: cn_jrc_es:
  • model: ConceptNet Numberbatch
  • corpus: JRC-Acquis
  • language: Spanish

• Labels:
  • <model>:
    cn - ConceptNet Numberbatch
    um - USE-ML
     
  • <dataset>
    aspec - ASPEC (Asian Scientific Paper Excerpt Corpus) - English and Japanese
    aspecx - ASPEC (Asian Scientific Paper Excerpt Corpus) - Japanese and Chinese
    jrc - JRC-Acquis
    europarl - Europarl
    pan - PAN-PC-11
     
  • <language>
    en - English
    es - Spanish
    fr - French
    ja - Japanese
    zh - Chinese